Real-time video communication is critically hampered by packet loss. Building on the neural video codec (NVC) and exploiting the statistics of its latent representations, we propose a channel importance-aware redundancy framework that eliminates insignificant channels and reallocates high redundancy to protect important channels for loss-resilient video coding. We first develop a channel-wise importance evaluation method based on the in-distribution (IND) region length criterion [1] for latent representations of prediction residuals and motion vectors produced by NVCs. Consequently, we derive a heuristic that adapts the number of protected channels to the packet loss rate (PLR). More channels are retained at low PLRs to improve video quality, while only a few critical channels are kept and assigned higher redundancy for enhanced robustness at high PLRs. Experimental results show that the proposed method provides more graceful quality degradation and maintains significant performance advantages, particularly under high PLRs, as illustrated in Figure 1.

Despite strong generalization ability, large language models (LLMs) are resource-intensive to deploy. Post-training quantization (PTQ) effectively compresses pretrained LLMs without retraining. Although recent studies target extremely low-bit quantization, most methods rely on extra parameters and the effective bit-width usually exceeds 2 bits. We introduce a saliency-guided mixed-precision PTQ framework that achieves an average bit-width of 1.70 bits while maintaining competitive performance. The proposed method allocates bit-widths across weight groups under a strict total budget, guided by a cross-precision saliency metric. The key innovation is a learnable affine transformation consisting of rotation, scaling, and shifting to align quantization grids with local weight distributions to minimize error. Experiments show that the proposed method matches state-of-the-art low-bit PTQ performance, while reducing the average bit-width by approximately 0.55 bits in Table 1.

Task-adapted compressed sensing magnetic resonance imaging (CS-MRI) is emerging to address the specific demands of downstream clinical tasks with significantly fewer k-space measurements than required by Nyquist sampling. However, existing task-adapted CS-MRI methods suffer from the uncertainty problem for medical diagnosis and cannot achieve adaptive sampling in end-to-end optimization with reconstruction or clinical tasks. To address these limitations, we propose the first task-adapted CS-MRI from the information-theoretic perspective to simultaneously achieve probabilistic inference for uncertainty prediction and adapt to arbitrary sampling ratios and versatile clinical applications. Specifically, we formalize the task-adapted CS-MRI optimization problem by maximizing the mutual information between undersampled k-space measurements and clinical tasks to enable probabilistic inference for addressing the uncertainty problem. We leverage amortized optimization and construct tractable variational bounds for mutual information to jointly optimize sampling, reconstruction, and task-inference models, which enables flexible sampling ratio control using a single end-to-end trained model. Furthermore, the proposed framework addresses two kinds of distinct clinical scenarios within a unified approach, i.e., i) joint task and reconstruction, where reconstruction serves as an auxiliary process to enhance task performance; and ii) task implementation with suppressed reconstruction, applicable for privacy protection. Extensive experiments on large-scale MRI datasets demonstrate that the proposed framework achieves highly competitive performance on standard metrics like Dice compared to deterministic counterpart but provides better distribution matching to the ground-truth posterior distribution as measured by the generalized energy distance (GED).

Pooling and unpooling are indispensable in constructing hierarchical spherical convolutional neural networks (HS-CNNs). Most existing models employ simple downsampling-based pooling, which ignores the sampling theorem and cannot adapt to different spherical signals (with different spectra) and tasks (dependent on different frequency components), thus suffering a significant information loss. Besides, signals reconstructed by the widely-adopted padding-based unpooling may also change unwantedly the spectra of original signals. To address these, we propose a novel framework of HS-CNNs with lifting structures to learn adaptive spherical wavelets for pooling and unpooling, named LiftHS-CNNs. Specifically, we learn spherical wavelets with a lifting structure to adaptively partition the input signal into low- and high-frequency sub-bands, with the down-scaled representations for pooling generated to preserve more information in the low-frequency sub-band. The lifting structure consists of learnable update and predict operators parameterized with graph attention to jointly consider the signal's characteristics and underlying geometries. We then propose an unpooling operation invertible to the lifting-based pooling for restoring the up-scaled representations, which can well preserve spectral characteristics of the original signal. Particular properties (i.e., spatial locality, vanishing moments, and stability) of the learned wavelets and the information preserving ability of the proposed pooling and unpooling are further studied. Experiments on benchmark spherical datasets for a wide range of tasks verify the superiority of our LiftHS-CNNs.

Adaptive bitrate (ABR) streaming is a popular technique used to improve the quality of experience (QoE) for users who watch videos online, which, for example, can provide a smoother video playback by dynamically adjusting the requested video quality with associated bitrate according to the constrained yet diverse network conditions. Recently, learning-based ABR algorithms have achieved a notable performance gain with lower inference overhead than the conventional heuristic or model-based baselines. However, their performance may degrade significantly in an unseen network environment with time-varying and heterogeneous throughput dynamics. For a better generalization, in this paper, we propose a meta-reinforcement learning (meta-RL)-based neural ABR algorithm that is able to quickly adapt its policy to these unseen throughput dynamics. Specifically, we propose a model-free system framework comprising an inference network and a policy network. The inference network infers distribution of the latent representation for underlying dynamics based on the recent throughout context, while the policy network is trained to quickly adapt to the changing throughout dynamics with the sampled latent representation. To effectively learn the inference network and meta-policy on mixed dynamics of the practical ABR scenarios, we further design a variational information bottleneck theory-based loss function for training the inference and policy networks, whose objective is to strike a trade-off between brevity of the latent representation and expressiveness of the meta-policy. We also derive a theoretically necessary condition for the bitrate versions that yield higher long-term QoE, based on which a dynamic action pruning strategy is further developed for practical implementation. This pruning strategy can not only prevent unsafe policy outputs in midst of unseen throughput dynamics, but may also reduce the computational complexity of model-based ABR algorithms. Finally, the meta-training and meta-adaptation procedures of our proposed algorithm are implemented across a range of throughput dynamics. The empirical evaluations on various datasets containing real-world network traces verify that our algorithm surpasses the state-of-the-art ABR algorithms, particularly in terms of the average chunk QoE and fast adaptation across out-of-distribution throughput traces.

Point cloud attribute compression is challenged by fitting the attribute signals living on irregular geometric structures. Existing methods cannot achieve compact multiscale representation for high-fidelity reconstruction using the handcrafted transforms or deep learning-based techniques. In this paper, we propose a novel geometry-aware lifting-based multiscale network via spatial-channel lifting scheme for point cloud attribute compression. The proposed network cascades geometry-aware spatial lifting to reduce spatial redundancy by adaptively capturing irregular geometric structures and progressive channel lifting to progressively reduce channel-wise redundancy in multiscale representation. Furthermore, we design the split, predict, and update operations for geometry-aware spatial lifting to fully exploit the geometry information representing irregular structures. We develop geometry-aware adaptive split to equally split input points with significance scores indicating their dependencies, and propose geometry-aware cross-attention filtering for the predict and update operations for decorrelation based on geometry information. To our best knowledge, this paper achieves the first lifting-based learned transform for point cloud compression that enjoys reversibility guarantees of multiscale representation to enhance rate-distortion performance. Experimental results show that the proposed framework achieves state-of-the-art performance on extensive point cloud datasets, and outperforms latest MPEG G-PCC standard and most recent deep learning based methods.

Neural Image Compression (NIC) has achieved superior compression performance by modeling images as implicit feature representations, yet its practical deployment is severely hindered by computational overhead. Recently, GaussianImage was proposed as a computationally efficient explicit image representation paradigm, which renders images from 2D Gaussians. However, it suffers from inferior compression performance relative to mainstream NIC frameworks, mainly due to low compressibility and representation capability of explicit Gaussian parameters. To this end, we propose Pixel-Aligned Generalized 2D Splatting (PA-G2DS) as a computation-efficient and compression-friendly image representation format. Specifically, we deploy a learnable rendering function with implicit coefficients to enhance the reconstructed image quality and improve the compression ratio over explicit Gaussian coefficients. Incorporating the proposed PA-G2DS as a computationally efficient decoder, we further develop a suite of image codecs optimized for either compression ratios or flexible deployment scenarios. Experiments prove that the proposed codecs could achieve 30ms compression latency and millisecond-level decompression latency, reducing the performance and efficiency gap between implicit and explicit image representation. Furthermore, the proposed codec opens potential applications for NIC such as JPEG-like sequential decompression and random-access during decompression.

Sparse representation via dedicated dictionary learning facilitates image restoration by simultaneously preserving significant representation and removing trivial perturbations. However, existing dictionary learning under strict ℓ0-norm based sparsity constraint lacks of convergence guarantees and suffers from the dilemma between block artifacts and degraded quality in restoring fine details, and could hamper the consequent deep unfolded networks (DUNs). To address these problems, in this paper, we propose the first ℓ0-regularized dictionary learning with a theoretical guarantee of global convergence that achieves globally consistent sparse approximation for image restoration. Specifically, we first formulate an ℓ0-norm based optimization for dictionary learning that simultaneously achieves sparse approximation of local patches and ensures global consistency between the recovered and observed images. Consequently, we develop ℓ0-regularized global-local dictionary learning (ℓ0-GLDL) that alternately achieves sparse code pursuit, local dictionary update, and global image reconstruction based on alternating proximal linearized method. We demonstrate in theory that ℓ0-GLDL converges globally to a critical point. Furthermore, we design a deep unfolded network by unrolling ℓ0 -GLDL into diverse subblocks to strictly realize alternating optimization for stable image restoration. Experimental results demonstrate that the proposed methods achieve state-of-the-art performance and outperform existing deep learning based methods in image restoration tasks such as image denoising, image deblurring, and image inpainting.

Image denoising remains a challenging problem due to the ill-posed nature of recovering clean images from noisy observations. Prior-based modeling plays a vital role in addressing this challenge. Low-rank priors effectively exploit non-local self-similarity by modeling correlations across similar images, but often over-smooth fine textures. In contrast, sparse priors, especially in the form of Convolutional Sparse Coding (CSC), excel at preserving high-frequency texture details, yet typically neglect the structure across images. Recent unfolded CSC networks have improved denoising performance by combining CSC with deep networks, but most of them rely solely on sparsity, overlooking the complementary low-rank structure information. To overcome these limitations, we propose UCSC-LR, the first interpretable unfolded CSC network that jointly leverages both sparse and low-rank priors within a unified architecture. UCSC-LR unrolls both the Iterative Shrinkage-Thresholding Algorithm (ISTA) and Singular Value Thresholding (SVT) to strictly implement alternating minimization, enabling simultaneous pursuit of sparse representations and low-rank reconstructions. A dedicated Fusion Net is introduced to adaptively integrate features from the two priors, allowing the model to preserve both texture and structural content. To further enhance performance on color image denoising, UCSC-LR incorporates a lightweight channel attention mechanism to capture inter-channel dependencies. With shared, pre-learned convolutional dictionaries and efficient parameterization, UCSC-LR achieves state-of-the-art performance with low model complexity. Extensive experiments on grayscale, blind, and color denoising tasks validate the effectiveness of the proposed method, consistently surpassing existing CSC-based and deep learning-based denoisers.

Existing 2D-to-3D human pose estimation (HPE) methods struggle with the occlusion issue by enriching information like temporal and visual cues in the lifting stage. In this paper, we argue that these methods ignore the limitation of the sparse skeleton 2D input representation, which fundamentally restricts the 2D-to-3D lifting and worsens the occlusion issue. To address these, we propose a novel two-stage generative densification method, named Hierarchical Pose AutoRegressive Transformer (HiPART), to generate hierarchical 2D dense poses from the original sparse 2D pose. Specifically, we first develop a multi-scale skeleton tokenization module to quantize the highly dense 2D pose into hierarchical tokens and propose a Skeleton-aware Alignment to strengthen token connections. We then develop a Hierarchical AutoRegressive Modeling scheme for hierarchical 2D pose generation. With generated hierarchical poses as inputs for 2D-to-3D lifting, the proposed method shows strong robustness in occluded scenarios and achieves state-of-the-art performance on the single-frame-based 3D HPE. Moreover, it outperforms numerous multi-frame methods while reducing parameter and computational complexity and can also complement them to further enhance performance and robustness.

Controllability of video generation has been recently concerned in addition to the quality of generated videos. The main challenge to controllable video generation is to synthesize videos based on user-specified instance spatial locations and movement trajectories. However, existing methods suffer from a dilemma between the resource consumption, generation quality, and user controllability. As an efficient alternative to prohibitive training-based video generation, existing zero-shot video generation methods cannot generate high-quality and motion-consistent videos under the control of layouts and movement trajectories. In this paper, we propose a novel zero-shot method named IM-Zero that ameliorates instance-level motion controllable video generation with enhanced control accuracy, motion consistency, and richness of details to address this problem. Specifically, we first present a motion generation stage that extracts motion and textural guidance from keyframe candidates from pre-trained grounded text-to-image model to generate the desired coarse motion video. Subsequently, we develop a video refinement stage that injects the motion priors of pre-trained text-to-video models and detail priors of pre-trained text-to-image models into the latents of coarse motion videos to further enhance video motion consistency and richness of details. To our best knowledge, IM-Zero is the first to simultaneously achieve high-quality video generation and allow to control both layouts and movement trajectories in a zero-shot manner. Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, inter-frame consistency, and the alignment of location and trajectory. Furthermore, compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances, finer control of specifying instance shapes via masks, and more difficult tasks of motion transfer for customizing fine-grained motion patterns through reference videos and high-quality text-to-video generation.

Autofocus is a crucial component of modern digital cameras. While recent learning-based methods achieve state-of-the-art in focus prediction accuracy, they unfortunately ignore the potential focus hunting phenomenon of back-and-forth lens movement in the multi-step focusing procedure. To address this, in this paper, we propose an expert regularized deep reinforcement learning (DRL)-based approach for autofocus, which can utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our method generally follows an actor-critic framework. To accelerate the DRL’s training with a higher sample efficiency, we initialize the policy with a pre-trained single-step prediction network, where the network is further improved by modifying the output of absolute in-focus position distribution to the relative lens movement distribution to establish a better mapping between input images and lens movement. To further stabilize DRL’s training with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network’s training. Empirical evaluations show that our method outperforms those learning-based methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate.

Parameter-efficient transfer learning (PETL) has emerged as a promising solution to adapt large-scale pre-trained models to downstream tasks. Nevertheless, these methods have not thoroughly explored the characteristics of PETL methods to optimize the fine-tuning performance with miminal volume of parameters. In this paper, we first reveal that, compared to pre-trained models, PETL tends to generate similar features via homogeneous feature transformations across different layers. Subsequently, we propose a Global Share Local Transform Mixture-of-Experts framework, namely GLEAM, that decomposes the adapter into a shared component and layer-specific local components to simultaneously reduce the redundancy in layer-wise parameter matrices for homogeneous feature transformations and fine-tune the locally specific parameters for minimizing performance loss. Specifically, we develop a shared mixture of convolution that introduces shared multi-scale sparse MoE to enable diverse transformations for suppressing the homogeneity issue of feature transformations in PETL. GLEAM is evaluated on more than 20 datasets for image classification and few-shot learning. Extensive experimental results demonstrate that it achieves comparable performance with existing PETL methods like LoRA with only 3% of its parameters and further yields competitive performance using only 0.07M parameters.

Generative diffusion models have shown empirical successes in point cloud resampling, generating a denser and more uniform distribution of points from sparse or noisy 3D point clouds by progressively refining noise into structure. However, existing diffusion models employ manually predefined schemes, which often fail to recover the underlying point cloud structure due to the rigid and disruptive nature of the geometric degradation. To address this issue, we propose a novel learnable heat diffusion framework for point cloud resampling, which directly parameterizes the marginal distribution for the forward process by learning the adaptive heat diffusion schedules and local filtering scales of the time-varying heat kernel, and consequently, generates an adaptive conditional prior for the reverse process. Unlike previous diffusion models with a fixed prior, the adaptive conditional prior selectively preserves geometric features of the point cloud by minimizing a refined variational lower bound, guiding the points to evolve towards the underlying surface during the reverse process. Extensive experimental results demonstrate that the proposed point cloud resampling achieves state-of-the-art performance in representative reconstruction tasks including point cloud denoising and upsampling.

Learned image compression (LIC) has demonstrated superior rate-distortion (R-D) performance compared to traditional codecs, but is challenged by training inefficiency that could incur more than two weeks to train a state-of-the-art model from scratch. Existing LIC methods overlook the slow convergence caused by compacting energy in learning nonlinear transforms. In this paper, we first reveal that such energy compaction consists of two components, \emph{i.e.}, feature decorrelation and uneven energy modulation. On such basis, we propose a linear auxiliary transform (AuxT) to disentangle energy compaction in training nonlinear transforms. The proposed AuxT obtains coarse approximation to achieve efficient energy compaction such that distribution fitting with the nonlinear transforms can be simplified to fine details. We then develop wavelet-based linear shortcuts (WLSs) for AuxT that leverages wavelet-based downsampling and orthogonal linear projection for feature decorrelation and subband-aware scaling for uneven energy modulation. AuxT is lightweight and plug-and-play to be integrated into diverse LIC models to address the slow convergence issue. Experimental results demonstrate that the proposed approach can accelerate training of LIC models by 2 times and simultaneously achieves an average 1\% BD-rate reduction. To our best knowledge, this is one of the first successful attempt that can significantly improve the convergence of LIC with comparable or superior rate-distortion performance.

In real-time communication (RTC) systems, a congestion control (CC) is indispensable to guarantee the end user’s quality of experience (QoE), the key of which is to predict the bottleneck link capacity and thus set the target bitrate for the sender. Existing deep reinforcement learning (DRL)-based bandwidth prediction methods, though improving the bandwidth prediction performance based on the pre-collected offline dataset, are usually designed for a single RTC scenario with poor generalization ability to other scenarios. What’s worse, the policy learned by these DRL methods can hardly capture the complex multi-modal distribution of underlying optimal policy. To address this, in this paper, we propose an offline DRL-based meta-diffusion policy, where we pre-train an encoder to extract the effective feature embedding of network dynamics, and further learn an expressive meta-policy based on the diffusion model to predict the bandwidth conditioned on the observed network states and extracted network dynamics features. Experiments demonstrate the enhanced accuracy and generalization of our meta-diffusion policy over existing landmark schemes, with a performance gain of 7.84% and 27.12%, in terms of the bandwidth prediction error rate and overestimate rate, respectively.

Federated learning (FL) is greatly challenged by the communication bottleneck and computation limitation on clients. Existing methods based on quantization for FL cannot simultaneously reduce the uplink and downlink communication cost and mitigate the computation burden on clients. To address this problem, in this paper, we propose the first low-bit integerized federated learning (LBI-FL) framework that quantizes the weights, activations, and gradients to lower than INT8 precision to evidently reduce the communication and computational costs. Specifically, we achieve dynamical temporal bit-width allocation for weights, activations, and gradients along the training trajectory via reinforcement learning. An agent is trained to determine bit-width allocation by comprehensively considering the states like current bit-width, training stage, and quantization loss as the state. The agent efficiently trained on small-scale datasets can be well generalized to train varying network architectures on non-independent and identically distributed datasets. Furthermore, we demonstrated in theory that federated learning with gradient quantization achieves an equivalent convergence rate to FedAvg. The proposed LBI-FL can reduce the communication costs by 8 times compared to full-precision FL. Extensive experiments show that the proposed LBI-FL achieves a reduction of more than 50\% BitOPs per client on average for FL with less than 2\% accuracy loss compared to low-bit training with INT8 precision.

The significant communication overhead and client data heterogeneity have posed an important challenge to current federated learning (FL) paradigm. Existing compression-based and optimization-based FL algorithms typically focus on addressing either the model compression challenge or the data heterogeneity issue individually, rather than tackling both of them. In this paper, we observe that by symbolizing the client model updates to be uploaded (i.e., normalizing the magnitude for each model parameter at local clients), the model heterogeneity, essentially stemmed from data heterogeneity, can be mitigated, and thereby helping improve the overall generalization performance of the globally aggregated model at the server. Inspired with this observation, and further motivated by the success of Lion optimizer in achieving the optimal performance on most tasks in the centralized learning, we propose a new FL algorithm, called FedSMU, which simultaneously reduces the communication overhead and alleviates the data heterogeneity issue. Specifically, FedSMU splits the standard Lion optimizer into the local updates and global execution, where only the symbol of client model updates commutes between the client and server. We theoretically prove the convergence of FedSMU for the general non-convex settings. Through extensive experimental evaluations on several benchmark datasets, we demonstrate that our FedSMU algorithm not only reduces the communication overhead, but also achieves a better generalization performance than the other compression-based and optimization-based baselines.

We propose Noise Conditional Variational Score Distillation (NCVSD), a novel method for distilling pretrained diffusion models into generative denoisers. We achieve this by revealing that the unconditional score function implicitly characterizes the score function of denoising posterior distributions. By integrating this insight into the Variational Score Distillation (VSD) framework, we enable scalable learning of generative denoisers capable of approximating samples from the denoising posterior distribution across a wide range of noise levels. The proposed generative denoisers exhibit desirable properties that allow fast generation while preserve the benefit of iterative refinement: (1) fast one-step generation through sampling from pure Gaussian noise at high noise levels; (2) improved sample quality by scaling the test-time compute with multi-step sampling; and (3) zero-shot probabilistic inference for flexible and controllable sampling. We evaluate NCVSD through extensive experiments, including class-conditional image generation and inverse problem solving. By scaling the test-time compute, our method outperforms teacher diffusion models and is on par with consistency models of larger sizes. Additionally, with significantly fewer NFEs than diffusion-based methods, we achieve record-breaking LPIPS on inverse problems.

Fourier phase retrieval (FPR) is a longstanding ill-posed non-convex inverse problem suffering from failure in reconstruction. Existing methods usually mitigate the issue in a trial-and-error fashion and cannot identify the failure to flexibly adjust the FPR process. In this paper, we propose FA-FPR, a novel failure aware method that guides FPR with diffusion priors by failure detection. We first discover that unsuccessful reconstructions can be identified as fake images due to their distinct structures from true signals, and then leverage a plug-and-play fake image detector to automatically determine the approximation to true signals. Furthermore, we develop a hybrid decoupling scheme that adaptively combines Gaussian noise and predicted noise by score function based on the failure detection. Extensive experiments under various oversampling ratios with noisy and noiseless measurements demonstrate that FA-FPR requires less than 50% number of repeated experiments on average to achieve the state-of-the-art performance.

Source-free domain adaptation (SFDA) is a challenging task that tackles domain shifts using only a pre-trained source model and unlabeled target data. Existing SFDA methods are restricted by the fundamental limitation of source-target domain discrepancy. Non-generation SFDA methods suffer from unreliable pseudo-labels in challenging scenarios with large domain discrepancies, while generation-based SFDA methods are evidently degraded due to enlarged domain discrepancies in creating pseudo-source data. To address this limitation, we propose a novel generation-based framework named Diffusion-Driven Progressive Target Manipulation (DPTM) that leverages unlabeled target data as references to reliably generate and progressively refine a pseudo-target domain for SFDA. Specifically, we divide the target samples into a trust set and a non-trust set based on the reliability of pseudo-labels to sufficiently and reliably exploit their information. For samples from the non-trust set, we develop a manipulation strategy to semantically transform them into the newly assigned categories, while simultaneously maintaining them in the target distribution via a latent diffusion model. Furthermore, we design a progressive refinement mechanism that progressively reduces the domain discrepancy between the pseudo-target domain and the real target domain via iterative refinement. Experimental results demonstrate that DPTM outperforms existing methods by a large margin and achieves state-of-the-art performance on four prevailing SFDA benchmark datasets with different scales. Remarkably, DPTM can significantly enhance the performance by up to 18.6\% in scenarios with large source-target gaps.

In low-latency live streaming (LLLS), the adaptive bitrate algorithm plays a critical role in optimizing encoding bitrates to meet the strict end-to-end latency requirement, which is mostly several seconds or less. Existing methods usually rely on a fixed preset latency target, limiting their generalization capabilities across the heterogeneous scenarios. To address this, we propose an offline meta-reinforcement learning-based bitrate decision framework for LLLS, which incorporates the expertise of multiple state-of-the-art LLLS algorithms from their collected experience and constructs a universal bitrate selection policy via the offline training. Specifically, we establish an RL-based policy that adaptively adjusts the network throughput measurements and selects the encoding bitrates for LLLS based on the network conditions. The policy’s performance is enhanced by improving the robustness of short-term throughput estimations. To leverage the expertise of current LLLS algorithms, we collect their decision-making trajectories, and directly train the policy on these offline trajectories with implicit Q-learning. In addition, a meta-RL paradigm is further adopted to learn a universal policy that performs uniformly well across the heterogeneous network conditions and varying target latencies. Experimental results demonstrate that, compared to the other baselines, the proposed method achieves at least a 5.8% performance gain in terms of the overall quality of experience (QoE), across the diverse target-latency scenarios in the real-world throughput traces.

Progressive coding adapts to reliable image and video transmission over unstable network with fluctuating bandwidth with truncatable bitstreams produced by layer-wise conditional entropy coding. However, recent learned progressive image and video compression methods solely leverage table lookup or real-time computation for conditional entropy coding, and suffer from dramatically increased storage and computational overheads due to the expansion of codewords. To address this problem, in this paper, we propose an efficient conditional entropy coding scheme for learned progressive coding that incorporates the fast lookup-table and low-memory real-time computation based on statistical distribution of the codewords. Specifically, we build a smaller lookup table that stores only the symbols with high occurrence frequency and achieve real-time computation of those symbols with low occurrence frequency. Furthermore, we seamlessly integrate the proposed scheme into existing learned image and video compression frameworks for enhanced coding efficiency. Experimental results show that the proposed scheme accelerates both encoding and decoding processes for learned progressive image and video compression. Remarkably, we for the first time successfully extend the proposed scheme to learned video compression to demonstrate its feasibility.

Cross-layer feature fusion critically influences the performance of deep neural networks, where mainstream approaches like additive fusion of residual connections struggle to model high-order nonlinear interactions, limiting representational capacity, while concatenative fusion of dense connections incurs significant memory and computational overhead. Although recent attention-based feature fusion methods refine feature representations and strengthen hierarchical interactions, they often struggle to balance expressiveness and efficiency. To address these limitations, we propose the Chebyshev Fusion Unit (CFU), a lightweight yet effective cross-layer fusion mechanism. Specifically, CFU computes high-order Chebyshev polynomial terms between residual features and current-layer features, each of which is aggregated through learnable scalar weights, forming enhanced fusion features. This design enables explicit modeling of complex cross-layer dependencies with minimal additional parameters. Extensive experiments across various commonly used base models on image classification, medical image segmentation, and physical law learning tasks demonstrate its superior performance.

Existing methods for integerized training speed up deep learning by using low-bitwidth integerized weights, activations, gradients, and optimizer buffers. However, they overlook the issue of full-precision latent weights, which consume excessive memory to accumulate gradient-based updates for optimizing the integerized weights. In this paper, we propose the first latent weight quantization schema for general integerized training, which minimizes quantization perturbation to training process via residual quantization with optimized dual quantizer. We leverage residual quantization to eliminate the correlation between latent weight and integerized weight for suppressing quantization noise. We further propose dual quantizer with optimal nonuniform codebook to avoid frozen weight and ensure statistically unbiased training trajectory as full-precision latent weight. The codebook is optimized to minimize the disturbance on weight update under importance guidance and achieved with a three-segment polyline approximation for hardware-friendly implementation. Extensive experiments show that the proposed schema allows integerized training with lowest 4-bit latent weight for various architectures including ResNets, MobileNetV2, and Transformers, and yields negligible performance loss in image classification and text generation. Furthermore, we successfully fine-tune Large Language Models with up to 13 billion parameters on one single GPU using the proposed schema.

Recent advances in self-supervised learning have witnessed great achievements, especially with the introduction of contrastive learning, where the goal is to maximize the mutual information between different augmentations of the same image, i.e., positive pairs. However, such optimization does not necessarily correspond to optimal representation due to noisy samples, thus inevitably being over-confident in the relevance between views. As a result, the learned model would capture spurious correlation and retain superfluous information that deteriorates representations. In this paper, we facilitate contrastive learning by reducing superfluous relevance between positive views. To this end, we introduce the representation entropy minimization regularization over the objective of vanilla contrastive learning, which forces representations to retain possibly the least information, thus alleviating superfluous relevance from irrelevant views. Then, we derive the analytical expression of the proposed objective by converting it to an information bottleneck problem and solving via variation approximation, which leads to a novel contrastive learning framework, termed as CLIMB, short for Contrastive Learning via variational InforMation Bottleneck. Experiments over multiple benchmarks demonstrate that CLIMB brings consistent improvement. Notably, using DINO as an instantiation, CLIMB achieves 4.5% and 3.5% gain under the k-NN classification metric with EfficientNet-B0 and ResNet-50 as backbones, respectively.

Federated learning (FL) commonly encourages the clients to perform multiple local updates before the global aggregation, thus avoiding frequent model exchanges and relieving the communication bottleneck between the server and clients. Though empirically effective, the negative impact of multiple local updates on the stability of FL is not thoroughly studied, which may result in a globally unstable and slow convergence. Based on sensitivity analysis, we define in this paper a local-update stability index for the general FL, as measured by the maximum inter-client model discrepancy after the multiple local updates that mainly stems from the data heterogeneity. It enables to determine how much the variation of client's models with multiple local updates may influence the global model, and can also be linked with the convergence and generalization. We theoretically derive the proposed local-update stability for current state-of-the-art FL methods, providing possible insight to understanding their motivation and limitation from a new perspective of stability. For example, naively executing the parallel acceleration locally at clients would harm the local-update stability. Motivated by this, we then propose a novel accelerated yet stabilized FL algorithm (named FedANAG) based on the server- and client-level Nesterov accelerated gradient (NAG). In FedANAG, the global and local momenta are elaborately designed and alternatively updated, while the stability of local update is enhanced with help of the global momentum. We prove the convergence of FedANAG for strongly convex, general convex and non-convex settings. We then conduct evaluations on both the synthetic and real-world datasets to first validate our proposed local-update stability. The results further show that across various data heterogeneity and client participation ratios, FedANAG not only accelerates the global convergence by reducing the required number of communication rounds to a target accuracy, but converges to an eventually higher accuracy.

End-to-end optimization via deep neural networks has facilitated lossy image compression. Existing neural network-based entropy models for end-to-end optimized image compression are limited by parameterized Gaussian distributions with deterministic mean and variance and cannot achieve accurate rate estimation for bottleneck representation with varying statistics. In this paper, we propose a novel entropy model based on deep Gaussian process regression (DGPR) to address this problem. Specifically, the proposed entropy model leverages autoregressive DGPR to flexibly predict the channel-wise posterior distributions of high-dimensional bottleneck representation for entropy coding. Consequently, we develop a well-established bit-rate estimation scheme via posterior inference of DGPR using the learned probabilistic distribution. Furthermore, scalable training is achieved via tensor train decomposition and Monte Carlo sampling to enable tractable variational inference of DGPR. To our best knowledge, this paper is the first attempt to develop the learnable probabilistic model for flexible parameter estimation in entropy modeling. Experimental results show that the proposed model outperforms conventional image compression methods (e.g., JPEG2000 and BPG) as well as recent end-to-end optimized methods on the Kodak and Tecnick datasets in terms of rate-distortion performance.

Entropy modeling is the core component of learned image compression (LIC) that models the distribution of latent representation learned from input images via neural networks for bit-rate estimation. However, existing entropy models employ presumed parameterized distributions such as Gaussian models and are limited for the learned latent representation characterized by complex distributions. To address this problem, in this paper, we for the first time achieve generative probabilistic entropy modeling of latent representation based on conditional diffusion models. Specifically, we propose a conditional diffusion-based probabilistic entropy model (CDPEM) to parameterize the latent representation with distributions of arbitrary forms that are generated by well designed training-test consistent denoising diffusion implicit model (TC-DDIM) without introducing any presumption. TC-DDIM is designed to leverage ancestral sampling to gradually approximate the distribution of latent representation with guaranteed consistency in generation for training and test. Furthermore, we develop a hierarchical spatial-channel context model to incorporate with TC-DDIM to sufficiently exploit spatial correlations with the approximate contextual information produced by ancestral sampling and channel-wise correlations using channel-wise information aggregation with reweighted training loss. Experimental results demonstrate that the proposed entropy model achieves state-of-the-art performance on the Kodak, CLIC, and Tecnick datasets compared to existing LIC methods. Remarkably, when incorporated with recent baselines, the proposed model outperforms latest VVC standard by an evident gain in R-D performance.

Variable-rate coding is challenging but indispensable for learned image compression (LIC) that is in nature characterized by nonlinear transform coding (NTC). Existing methods for variable-rate LIC are restricted by the non-smooth quantization process with zero gradients almost everywhere, and consequently, suffer from training-test gap and degraded rate-distortion (R-D) performance. To address this problem, in this paper, we propose sampling-based optimization for training NTC models along with non-uniform quantizers. Different from gradient-based optimization, the proposed sampling-based optimization first randomly samples the parameters from Gaussian distributions with progressively reduced variance and then selects the optimal parameters with a R-D indicator. On the basis of sampling-based optimization, we develop a learnable non-uniform dead-zone quantizer by adaptively refining the quantization steps for variable-rate coding with nonlinear transforms. Furthermore, we incorporate the learnable dead-zone quantizer to achieve a variable-rate LIC model with enhanced R-D performance and design rate and distortion control algorithms to adapt to dynamic network conditions. Experimental results show that the proposed method achieves state-of-the-art R-D performance in variable-rate image compression. It obtains an average 8.82% BD-rate reduction compared to latest versatile video compression (VVC) standard, and simultaneously achieves precise rate and distortion control with an average variation of 0.0087 bpp in bit-rates and 0.1265 dB in distortion on the Kodak dataset.

Image compression for both human and machine vision has become prevailing to accommodate to rising demands for machine-machine and human-machine communications. Scalable human-machine image compression is recently emerging as an efficient alternative to simultaneously achieve high accuracy for machine vision in the base layer and obtain high-fidelity reconstruction for human vision in the enhancement layer. However, existing methods achieve scalable coding with heuristic mechanisms, which cannot fully exploit the inter-layer correlations and evidently sacrifice rate-distortion performance. In this paper, we propose task-adapted learnable embedded quantization to address this problem in an analytically optimized fashion. We first reveal the relationship between the latent representations for machine and human vision and demonstrate that optimal representation for machine vision can be approximated with post-training optimization on the learned representation for human vision. On such basis, we propose task-adapted learnable embedded quantization that leverages learnable step predictor to adaptively determine the optimal quantization step for diverse machine vision tasks such that inter-layer correlations between representations for human and machine vision are sufficiently exploited using embedded quantization. Furthermore, we develop a human-machine scalable coding framework by incorporating the proposed embedded quantization into pre-trained learned image compression models. Experimental results demonstrate that the proposed framework achieves state-of-the-art performance on machine vision tasks like object detection, instance segmentation, and panoptic segmentation with negligible loss in rate-distortion performance for human vision.

Existing pretraining methods for semantic segmentation are hampered by the task gap between global image -level pretraining and local pixel-level finetuning. Joint dense-level pretraining is a promising alternative to exploit off-the-shelf annotations from diverse segmentation datasets but suffers from low-quality class embeddings and inconsistent data and supervision signals across multiple datasets by directly employing CLIP. To overcome these challenges, we propose a novelMulti-datasEt harmoNized pretraining framework forSemantic sEgmentation (MENSA). MENSA incorporates high-quality language embeddings and momentum-updated visual embeddings to effectively model the class relationships in the embedding space and thereby provide reliable supervision information for each category. To further adapt to multiple datasets, we achieve one-to-many pixel-embedding pairing with cross-dataset multi-label mapping through cross-modal information exchange to mitigate inconsistent supervision signals and introduce region-level and pixel-level cross-dataset mixing for varying data distribution. Experimental results demonstrate that MENSA is a powerful foundation segmentation model that consistently outperforms popular supervised or unsupervised ImageNet pretrained models for various benchmarks under standard fine-tuning. Furthermore, MENSA is shown to significantly benefit frozen-backbone fine-tuning and zero-shot learning by endowing pixel-level distinctiveness to learned representations.

In federated learning (FL), the heterogeneity of data and asynchronous participation of clients have been observed to induce the local client's model discrepancy with high variance, leading to a slow and unstable convergence globally at the server. In this article, motivated by the usefulness of stale client updates, we first propose a general framework, named FedVR, to address this issue. In FedVR, we design an aggregate of both fresh and stale local model updates without additional communication overhead, which is computed at the server as a control variate to reduce the client variance incurred by data heterogeneity and client sampling. In order to further reduce the model discrepancy between local clients, we therefore propose FedMDVR, which broadcasts the designed control variate to all the active clients to help correct their local update directions toward the global optimum, i.e., stationary point of the global objective function. While in the global update at server, the client variance is also decreased as inherited from the variance reduction nature of FedVR. We theoretically prove the convergence of FedVR and FedMDVR in the general nonconvex settings. Through extensive experimental evaluations on several benchmark datasets, we also demonstrate that our proposed FedVR and FedMDVR not only accelerate the convergence by reducing the number of communication rounds required to achieve a certain target accuracy, but more importantly, can converge to a higher accuracy than the baseline algorithms.

Network planning is crucial to facilitate network service under limited network operation costs. However, adapting the network topology (i.e., connections and capacities for physical and IP links) to time-varying and stochastic network states is challenging due to high computational overhead of problem solving. Existing deep learning based methods suffer from prohibitive computational complexity due to simultaneously solving IP link capacities and allocated traffics or iteratively exploring search space with excessive samples via heuristic rules. In this paper, we propose GDPlan, the first generative framework that leverages conditional graph diffusion model to address this challenge. To achieve high solving efficiency, GDPlan decouples network planning into two separate stages, i.e., generation of diverse high-quality solutions to IP link capacity and refinement of the solutions under feasibility constraints. Specifically, we develop generation-oriented graph signal modeling that reformulates solving IP link capacities as generation of graph topology conditioned on the graph signals related to traffic demands, physical links and network costs. Consequently, energy-guided controllable graph generation is achieved to capture significant structural patterns of graph topology with score-based graph diffusion model and produce diverse solutions with a guarantee of feasibility under specific objectives or constraints. GDPlan successfully achieves graph generation based solution to network planning with an order of magnitude higher solving speed than the commercial solver Gurobi. Experimental results demonstrate that, compared with Gurobi, the proposed GDPlan obtains lowest 3.7% average gap with only about 7.5% average running time.

Learned image compression (LIC) has recently emerged as a promising alternative to realize transform-based coding. Contrary to widely studied transform and entropy modeling in LIC, most LIC methods are limited to uniform scalar quantization and existing vector quantization (VQ) methods are restricted by the soft-to-hard quantization strategy. In this paper, we propose a novel entropy relaxed lattice vector quantization (ERLVQ) scheme for LIC. ERLVQ achieves a tractable dimension-wise approximation to the probability mass function for the complex Voronoi regions in a general class of lattices. Instead of directly performing probabilistic estimation on the lattices, ERLVQ equivalently aligns complex Voronoi regions with a set of complete and non-overlapping hypercubes in a projected parametric domain. Space covering technique is then developed to generating hypercubes with their centers fixed to integer projected lattice points and their volumes equaling the Voronoi regions to enable separable dimension-wise estimation of multi-variate probability mass functions. The proposed ERLVQ is the first to adapt to arbitrary lattices for VQ based on LIC and can be embedded into existing LIC methods. Experimental results demonstrate that the proposed ERLVQ is superior to the uniform scalar quantization (USQ) and existing lattice VQ methods [1] on prevailing LIC baselines, as shown in Table 1 .

Network quantization is promising to alleviate the storage and deployment burden for deep neural networks. However, existing methods focus on directly quantizing the network parameters but ignore the correlations between the initialized parameters (generated by random seeds) and the optimized parameters (obtained when the network training converges). In this paper, we reveal the redundancy between the initialized and optimized parameters and propose a novel quantization scheme for network parameters with rescaled initialization. We develop a novel quantization scheme for quantizing the delta parameters, i.e. , the difference between the optimized parameters and the rescaled initialization, as shown in Figure 1 . The scale is learned to adapt the weight regularization in training by minimizing the rounding error and penalizing the long-tailed distribution of the delta parameter. Adaptive rounding is further presented to calibrate the rounding direction of the delta parameter via a learnable bias. The proposed method can achieve high-efficiency prediction of network parameters and consistent performance gains with a trivial overhead of single random seed. Extensive experimental results demonstrate its superior performance on various popular network architectures, as exemplified in Table 1 .

State-of-the-art learning-based stability control methods for nonlinear robotic systems suffer from the issue of reality gap, which stems from discrepancy of the system dynamics between training and target (test) environments. To mitigate this gap, we propose an adversarially robust neural Lyapunov control (ARNLC) method to improve the robustness and generalization capabilities for Lyapunov theory-based stability control. Specifically, inspired by adversarial learning, we introduce an adversary to simulate the dynamics discrepancy, which is learned through deep reinforcement learning to generate the worst-case perturbations during the controller’s training. By alternatively updating the controller to minimize the perturbed Lyapunov risk and the adversary to deviate the controller from its objective, the learned control policy enjoys a theoretical guarantee of stability. Empirical evaluations on five stability control tasks with the uniform and worst-case perturbations demonstrate that ARNLC not only accelerates the convergence to asymptotic stability, but can generalize better in the entire perturbation space.

Existing diffusion models for inverse problems have demonstrated impressive performance but suffer from prohibitive sampling complexity due to lengthy iterative sampling procedures. While introducing pre-trained consistency models (CMs) as priors holds promise for fast and high-quality sampling, theoretical disparities between CMs and diffusion models remain, hindering the application of CMs in solving inverse problems. To address this issue, we propose a novel framework, named Dual Data Consistency Guided Consistency Model (DCCM), that for the first time to solve inverse problems with pretrained CM priors. We establish a denoising interpretation of CMs to set up the equivalence between CMs and denoisers and incorporate CM in a theoretically sound fashion. Consequently, we develop refined data consistency to facilitate optimization with CM priors and avoid local minima caused by the nonlinearity of degradation operators. Furthermore, we introduce the data consistency shortcut that leverages the manifold hypothesis to approximate refined data consistency and bypass backpropagation for enhanced sampling speed without reconstruction quality loss. Extensive experiments demonstrate DCCM achieves state-of-the-art performance in terms of reconstruction quality and sampling speed in a wide range tasks of image deblurring, super-resolution, and inpainting.

Learned image compression (LIC) has gained traction as an effective solution for image storage and transmission in recent years. However, existing LIC methods are redundant in latent representation due to limitations in capturing anisotropic frequency components and preserving directional details. To overcome these challenges, we propose a novel frequency-aware transformer (FAT) block that for the first time achieves multiscale directional ananlysis for LIC. The FAT block comprises frequency-decomposition window attention (FDWA) modules to capture multiscale and directional frequency components of natural images. Additionally, we introduce frequency-modulation feed-forward network (FMFFN) to adaptively modulate different frequency components, improving rate-distortion performance. Furthermore, we present a transformer-based channel-wise autoregressive (T-CA) model that effectively exploits channel dependencies. Experiments show that our method achieves state-of-the-art rate-distortion performance compared to existing LIC methods, and evidently outperforms latest standardized codec VTM-12.1 by 14.5\%, 15.1\%, 13.0\% in BD-rate on the Kodak, Tecnick, and CLIC datasets.

As two prominent strategies for representation learning, Contrastive Learning (CL) and Masked Image Modeling (MIM) have witnessed significant progress. Previous studies have demonstrated the advantages of each approach in specific scenarios. CL, resembling supervised pre-training, excels at capturing longer-range global patterns and enhancing feature discrimination, while MIM is adept at introducing local and diverse attention across transformer layers. Considering the respective strengths, previous studies utilize feature distillation to inherit both discrimination and diversity. In this paper, we thoroughly examine previous feature distillation methods and observe that the increase in diversity mainly stems from asymmetric designs, which may in turn compromise the discrimination ability. To strike a balance between the two properties, we propose a simple yet effective strategy termed Hybrid Distill, which leverages both the CL and MIM teachers to jointly guide the student model. Hybrid Distill emulates the token relations of the MIM teacher at intermediate layers for diversity, while simultaneously distilling the final features of the CL teacher to enhance discrimination. A progressive redundant token masking strategy is employed to reduce the expenses associated with distillation and aid in preventing the model from converging to local optima. Experimental results demonstrate that Hybrid Distill achieves superior performance on various benchmark datasets.

Pre-training followed by full fine-tuning has gradually been substituted by Parameter-Efficient Tuning (PET) in the field of computer vision. PET has gained popularity, especially in the context of large-scale models, due to its ability to reduce transfer learning costs and conserve hardware resources. However, existing PET approaches primarily focus on recognition tasks and typically support uni-modal optimization, while neglecting dense prediction tasks and vision language interactions. To address this limitation, we propose a novel PET framework called **B**i-direction**a**l Inte**r**twined Vision **L**anguage Effici**e**nt Tuning for **R**eferring **I**mage Segment**a**tion (**BarLeRIa**), which leverages bi-directional intertwined vision language adapters to fully exploit the frozen pre-trained models' potential in cross-modal dense prediction tasks. In BarLeRIa, two different tuning modules are employed for efficient attention, one for global, and the other for local, along with an intertwined vision language tuning module for efficient modal fusion. Extensive experiments conducted on RIS benchmarks demonstrate the superiority of BarLeRIa over prior PET methods with a significant margin, i.e., achieving an average improvement of 5.6\%. Remarkably, without requiring additional training datasets, BarLeRIa even surpasses SOTA full fine-tuning approaches. The code is available at https://github.com/NastrondAd/BarLeRIa.

Low-bit integer training emerges as a promising approach to mitigate the heavy burden during network training by quantizing the weights, activations, and gradients. However, existing methods cannot well achieve mixed-precision quantization for low-bit training and are commonly limited to INT8 precision. In this paper, we propose a novel low-bit integer training framework that, for the first time, achieves adaptive mixed-precision allocation (AMPA) for weights, activations, and gradients, and pushes the boundaries to a precision level below INT8. We develop a novel magnitude-based sensitivity measurement with regard to the quantization losses of weight, activation, and gradient quantization and the average gradient magnitudes, which is demonstrated as an upper bound of quantization influence in theory. We further design a layer-wise precision update strategy under observations on the quantization losses and their effects on model performance in low-bit training. Extensive experiments on different backbones and datasets show that, compared to INT8 quantization, the proposed method can achieve more than 38% BitOPs reduction with a tolerable loss below 2% in image classification, image segmentation, and language modeling.

Recent diffusion models provide a promising zero-shot solution to noisy linear inverse problems without retraining for specific inverse problems. In this paper, we reveal that recent methods can be uniformly interpreted as employing a Gaussian approximation with hand-crafted isotropic covariance for the intractable denoising posterior to approximate the conditional posterior mean. Inspired by this finding, we propose to improve recent methods by using more principled covariance determined by maximum likelihood estimation. To achieve posterior covariance optimization without retraining, we provide general plug-and-play solutions based on two approaches specifically designed for leveraging pre-trained models with and without reverse covariance. We further propose a scalable method for learning posterior covariance prediction based on representation with orthonormal basis. Experimental results demonstrate that the proposed methods significantly enhance reconstruction performance without requiring hyperparameter tuning.

Existing self-supervised methods for gaze estimation using the dominant streams of contrastive and generative approaches are restricted to eye images and could fail in general full-face settings. In this paper, we reveal that contrastive methods are ineffective in data augmentation for self-supervised full-face gaze estimation, while generative methods are prone to trivial solutions due to the absence of explicit regularization on semantic representations. To address this challenge, we propose a novel approach called **B**ootstrap auto-**e**ncoders with **C**ontrastive p**a**radigm (**BeCa**), which combines the strengths of both generative and contrastive methods. Specifically, we revisit the Auto-Encoder used in generative approaches and incorporate the contrastive paradigm to introduce explicit regularization on gaze representation. Furthermore, we design the InfoMSE loss as an alternative to the vanilla MSE loss for Auto-Encoder to mitigate the inconsistency between reconstruction and representation learning. Experimental results demonstrate that the proposed approaches outperform state-of-the-art unsupervised gaze approaches on extensive datasets (including wild scenes) under both within-dataset and cross-dataset protocols.

Point cloud upsampling is crucial for 3D reconstruction, with recent research significantly benefitting from the advances in deep learning technologies. The majority of existing methods, which focus on a sequence of processes including feature extraction, augmentation, and the reconstruction of coordinates, encounter significant challenges in interpreting the geometric attributes they uncover, particularly with respect to the intricacies of transitioning feature dimensionality. In this paper, we delve deeper into modeling Partial Differential Equations (PDEs) specifically tailored for the inverse heat dissipation process in dense point clouds. Our goal is to detect gradients within the dense point cloud data distribution and refine the accuracy of interpolated points' positions along with their complex geometric nuances through a systematic iterative approximation method. Simultaneously, we adopt multivectors from geometric algebra as the primary tool for representing the geometric characteristics of point clouds, moving beyond the conventional vector space representations. The use of geometric products of multivectors enables us to capture the complex relationships between scalars, vectors, and their components more effectively. This methodology not only offers a robust framework for depicting the geometric features of point clouds but also enhances our modeling capabilities for inverse heat dissipation PDEs. Through both qualitative and quantitative assessments, we demonstrate that our results significantly outperform existing state-of-the-art techniques in terms of widely recognized point cloud evaluation metrics and 3D visual reconstruction fidelity.

Image compression for machine vision exhibits various rate-accuracy performance across different downstream tasks and content types. An efficient utilization of constrained network resource for achieving an optimal overall task performance has thus recently attracted a growing attention. In this paper, we propose Tombo, a task-oriented image compression and transmission framework that efficiently identifies the optimal encoding bitrate and routing scheme for multiple image bitstreams delivered simultaneously for different downstream tasks. Specifically, we study the characteristics of image rate-accuracy performance for different machine vision tasks, and formulate the task-oriented joint bitrate and routing optimization problem for multi-bitstreams as a multi-commodity network flow problem with the time-expanded network modeling. To ensure consistency between the encoding bitrate and routing optimization, we also propose an augmented network that incorporates the encoding bitrate variables into the routing variables. To improve computational efficiency, we further convert the original optimization problem to a multi-marginal optimal transport problem, and adopt a Sinkhorn iteration-based algorithm to quickly obtain the near-optimal solution. Finally, we adapt Tombo to efficiently deal with the dynamic network scenario where link capacities may fluctuate over time. Empirical evaluations on three typical machine vision tasks and four real-world network topologies demonstrate that Tombo achieves a comparable performance to the optimal one solved by the off-the-shelf solver Gurobi, with a 5x ~ 114× speedup.

Most of existing federated learning (FL) formulation is treated as a point-estimate of models, inherently prone to overfitting on scarce client-side data with overconfident decisions. Though Bayesian inference can alleviate this issue, a direct posterior inference at clients may result in biased local posterior estimates due to data heterogeneity, leading to a sub-optimal global posterior. From an information-theoretic perspective, we propose FedMDMI, a federated posterior inference framework based on model-data mutual information (MI). Specifically, a global model-data MI term is introduced as regularization to enforce the global model to learn essential information from the heterogeneous local data, alleviating the bias caused by data heterogeneity and hence enhancing generalization. To make this global MI tractable, we decompose it into local MI terms at the clients, converting the global objective with MI regularization into several locally optimizable objectives based on local data. For these local objectives, we further show that the optimal local posterior is a Gibbs posterior, which can be efficiently sampled with stochastic gradient Langevin dynamics methods. Finally, at the server, we approximate sampling from the global Gibbs posterior by simply averaging samples from the local posteriors. Theoretical analysis provides a generalization bound for FL w.r.t. the model-data MI, which, at different levels of regularization, represents a federated version of the bias-variance trade-off. Experimental results demonstrate a better generalization behavior with better calibrated uncertainty estimates of FedMDMI.

Diffusion Transformer (DiT) is emerging as a cutting-edge trend in the landscape of generative diffusion models for image generation. Recently, masked-reconstruction strategies have been considered to improve the efficiency and semantic consistency in training DiT but suffer from deficiency in contextual information extraction. In this paper, we provide a new insight to reveal that noisy-to-noisy masked-reconstruction harms sufficient utilization of contextual information. We further demonstrate the insight with theoretical analysis and empirical study on the mutual information between unmasked and masked patches. Guided by such insight, we propose a novel training paradigm named MC-DiT for fully learning contextual information via diffusion denoising at different noise variances with clean-to-clean mask-reconstruction. Moreover, to avoid model collapse, we design two complementary branches of DiT decoders for enhancing the use of noisy patches and mitigating excessive reliance on clean patches in reconstruction. Extensive experimental results on 256$\times$256 and 512$\times$512 image generation on the ImageNet dataset demonstrate that the proposed MC-DiT achieves state-of-the-art performance in unconditional and conditional image generation with enhanced convergence speed.

Autofocus (AF) is a fundamental task in digital image capturing, yet current approaches often exhibit a poor performance in terms of speed or accuracy. In this paper, we propose a deep reinforcement learning (DRL)-based framework for camera AF, by incorporating a pre-trained MobileNet-v2 encoder for image feature extraction and a Gaussian process regression (GPR) module for in-focus prediction. Specially, we formulate AF as a multi-step process, in which the lens movement trajectory is dynamically generated from a DRL-based policy. By leveraging image features extracted from the MobileNet-v2 encoder, the DRL-based policy can efficiently get close to the in-focus position, which enables the focus search in a fast speed, thus overcoming the drawback of slow focusing procedure suffered by traditional search-based AF methods. A GPR-based in-focus predictor is further designed to terminate the DRL’s exploration once it is able to provide a reliable prediction on the in-focus position, by exploiting the sharpness metric values of image patches captured at the DRL-generated trajectory of lens positions. In further contrast with the one-step prediction-based methods, we leverage the sequential information brought additionally by the multi-step exploration, enabling to gain a higher prediction accuracy. Experimental results demonstrate that our approach provides a significant improvement compared to both the search-based and prediction-based AF baselines.

By introducing randomness on the environments, domain randomization (DR) imposes diversity to the policy training of deep reinforcement learning, and thus improves its capability of generalization. The randomization of environments, however, introduces another source of variability for the estimate of policy gradients, in addition to the already high variance incurred by trajectory sampling. Therefore, with standard state-dependent baselines, the policy gradient methods may still suffer high variance, causing a low sample efficiency during the training of DR. In this paper, we theoretically derive a bias-free and state/environment-dependent optimal baseline for DR, and analytically show its ability to achieve further variance reduction over the standard constant and state-dependent baselines for DR. Based on our theory, we further propose a variance reduced domain randomization (VRDR) approach for policy gradient methods, to strike a tradeoff between the variance reduction and computational complexity for the practical implementation. By dividing the entire space of environments into some subspaces and then estimating the state/subspace-dependent baseline, VRDR enjoys a theoretical guarantee of variance reduction and faster convergence than the state-dependent baselines. Empirical evaluations on six robot control tasks with randomized dynamics demonstrate that VRDR not only accelerates the convergence of policy training, but can consistently achieve a better eventual policy with improved training stability.

3-D point clouds facilitate 3-D visual applications with detailed information of objects and scenes but bring about enormous challenges to design efficient compression technologies. The irregular signal statistics and high-order geometric structures of 3-D point clouds cannot be fully exploited by existing sparse representation and deep learning based point cloud attribute compression schemes and graph dictionary learning paradigms. In this paper, we propose a novel p-Laplacian embedding graph dictionary learning framework that jointly exploits the varying signal statistics and high-order geometric structures for 3-D point cloud attribute compression. The proposed framework formulates a nonconvex minimization constrained by p-Laplacian embedding regularization to learn a graph dictionary varying smoothly along the high-order geometric structures. An efficient alternating optimization paradigm is developed by harnessing ADMM to solve the nonconvex minimization. To our best knowledge, this paper proposes the first graph dictionary learning framework for point cloud compression. Furthermore, we devise an efficient layered compression scheme that integrates the proposed framework to exploit the correlations of 3-D point clouds in a structured fashion. Experimental results demonstrate that the proposed framework is superior to state-of-the-art transform-based methods in M-term approximation and point cloud attribute compression and outperforms recent MPEG G-PCC reference software.

Source-free domain adaptation (SFDA) shows the potential to improve the generalizability of deep learning-based face anti-spoofing (FAS) while preserving the privacy and security of sensitive human faces. However, existing SFDA methods are significantly degraded without accessing source data due to the inability to mitigate domain and identity bias in FAS. In this paper, we propose a novel Source-free Domain Adaptation framework for FAS (SDA-FAS) that systematically addresses the challenges of source model pre-training, source knowledge adaptation, and target data exploration under the source-free setting. Specifically, we develop a generalized method for source model pre-training that leverages a causality-inspired PatchMix data augmentation to diminish domain bias and designs the patch-wise contrastive loss to alleviate identity bias. For source knowledge adaptation, we propose a contrastive domain alignment module to align conditional distribution across domains with a theoretical equivalence to adaptation based on source data. Furthermore, target data exploration is achieved via self-supervised learning with patch shuffle augmentation to identify unseen attack types, which is ignored in existing SFDA methods. To our best knowledge, this paper provides the first full-stack privacy-preserving framework to address the generalization problem in FAS. Extensive experiments on nineteen cross-dataset scenarios show our framework considerably outperforms state-of-the-art methods.

Recent learned image compression models surpass manually designed methods in rate-distortion performance by introducing nonlinear transforms and end-to-end optimization. However, there still lack quantitative measurements that efficiently evaluate the latent representations inferred by learned image compression models. To address this problem, we develop novel measurements on robustness and importance of the latent representations. We first propose an admissible range that can be efficiently estimated via gradient ascent and descent for establishing the empirical distribution of latent representations. Consequently, the in-distribution region within the admissible range is derived to measure the robustness and channel importance of latent representations of natural images. Visualization demonstrates the statistics of latent representations are significantly distinguishing in the properties of robustness and linearity within and outside the in-distribution region. To our best knowledge, this paper proposes the first statistically meaningful measurements for learned image compression and successfully applies the measurements in corruption alleviation during successive image compression and post-training pruning in a training-free fashion. Compared with existing methods, the shrunk in-distribution constraint derived from the in-distribution region achieves superior robustness and rate-distortion performance in successive compression. The channel importance allows post-training pruning to achieve comparable rate-distortion performance with a reduction of up to 60% entropy coding time.

Predicting the saliency map of a 360-degree video is the key for various downstream tasks, such as saliency-based compression and tile-based adaptive streaming. Besides static salient objects, the moving target will also contribute to the saliency map. Therefore, the joint exploitation of spherical spatio-temporal information is necessary for an accurate saliency prediction. The spherical spatial feature extraction, however, is hindered by the non-Euclidean geometric nature of spherical data, which imposes difficulty on direct extraction of the spatial features with traditional convolutional neural networks (CNNs). While the efficient exploitation of temporal correlation between these spherical spatial features remains another challenge, which requires the extraction of spherical optical flows for explicit motion information. To address these, in this paper, we first propose a spherical graph-based Farneback algorithm to extract the spherical optical flows directly in the sphere domain, by leveraging the GICOPix uniform sampling scheme. We then design a 3D separable graph convolutional network-based saliency prediction framework, named 360Spred, by taking both the spherical frames and spherical optical flows as input. The proposed 360Spred framework is based on the U-Net structure, with a 3D separable graph convolution (3DSGC) operator that directly extracts the visual and motion features in the sphere domain and exploits temporal correlation of both the high-level and low-level spatial features. Experimental results on two public datasets show that 360Spred can achieve a better performance than other baseline models in terms of the saliency prediction accuracy for 360-degree videos.

In adaptive video streaming, the design of an adaptive bitrate (ABR) strategy is critical for the quality-of-experience (QoE) perceived by users. Though current learning-based ABR algorithms achieve state-of-the-art performance for users with a given QoE metric setting for training, they may unfortunately suffer the poor generalization issue for other users with different QoE preferences. Besides, how to quantitatively characterize the distinct QoE preference for a user has also not been extensively studied yet. In this paper, we propose STEER, a successor feature-based transfer reinforcement learning framework for fast learning the ABR strategies on heterogeneous QoE preferences. Specifically, we first develop a QoE preference analysis scheme to infer the personal QoE preference of a single user based on the user's actual viewing history. We then formulate the personalized QoE maximization problem as a reinforcement learning (RL) task, which optimizes the ABR strategy to maximize the overall QoE perceived by the user. Further, we model the QoE maximization problem for multiple users with heterogeneous QoE preferences as a multi-task RL problem, with each task distinguished by the user-distinct QoE preference. To efficiently address this problem, the proposed STEER solves for each RL-based ABR task by learning its optimal successor feature (SF) function, which can be exploited as shared knowledge across tasks to facilitate the transfer between tasks. With SF functions, STEER can quickly evaluate the optimal policies of previously learned tasks on a new task, and further use the generalized policy improvement operation to obtain a jumpstart policy. Both theoretically and empirically, we show that this jumpstart policy is a good initial policy with a performance guarantee for better generalization in the new task, and can also lead to a faster convergence to the optimal policy of the new task.

Viewers of 360-degree videos are provided with both visual modality to characterize their surrounding views and audio modality to indicate the sound direction. Though both modalities are important for saliency prediction, little work has been done by jointly exploiting them, which is mainly due to the lack of audio-visual saliency datasets and insufficient exploitation of the multi-modality. In this article, we first construct an audio-visual saliency dataset with 57 360-degree videos watched by 63 viewers. Through a deep analysis of the constructed dataset, we find that the human gaze can be attracted by the auditory cues, resulting in a more concentrated saliency map if the sound source's location is further provided. To jointly exploit the visual and audio features and their correlation, we further design a saliency prediction network for 360-degree videos (SVGC-AVA) based on spherical vector-based graph convolution and audio-visual attention. The proposed spherical vector-based graph convolution can process visual and audio features directly in the sphere domain, thus avoiding projection distortion incurred by traditional CNN-based predictors. In addition, the audio-visual attention scheme explores self-modal and cross-modal correlation for both modalities, which are further hierarchically processed with the U-Net's multi-scale structure of SVGC-AVA. Evaluations on both our and public datasets validate that SVGC-AVA can achieve higher prediction accuracy, both qualitatively and subjectively.

Message passing has evolved as an effective tool for designing graph neural networks (GNNs). However, most existing methods for message passing simply sum or average all the neighboring features to update node representations. They are restricted by two problems: 1) lack of interpretability to identify node features significant to the prediction of GNNs and 2) feature overmixing that leads to the oversmoothing issue in capturing long-range dependencies and inability to handle graphs under heterophily or low homophily. In this article, we propose a node-level capsule graph neural network (NCGNN) to address these problems with an improved message passing scheme. Specifically, NCGNN represents nodes as groups of node-level capsules, in which each capsule extracts distinctive features of its corresponding node. For each node-level capsule, a novel dynamic routing procedure is developed to adaptively select appropriate capsules for aggregation from a subgraph identified by the designed graph filter. NCGNN aggregates only the advantageous capsules and restrains irrelevant messages to avoid overmixing features of interacting nodes. Therefore, it can relieve the oversmoothing issue and learn effective node representations over graphs with homophily or heterophily. Furthermore, our proposed message passing scheme is inherently interpretable and exempt from complex post hoc explanations, as the graph filter and the dynamic routing procedure identify a subset of node features that are most significant to the model prediction from the extracted subgraph. Extensive experiments on synthetic as well as real-world graphs demonstrate that NCGNN can well address the oversmoothing issue and produce better node representations for semisupervised node classification. It outperforms the state of the arts under both homophily and heterophily.

The scene graph is a novel data structure describing objects and their pairwise relationship within image scenes. As the size of scene graphs in vision and multimedia applications increases, the need for lossless storage and transmission of such data becomes more critical. However, the compression of scene graphs is less studied because of the complicated data structures involved and complex distributions. Existing solutions usually involve general-purpose compressors or graph structure compression methods, which are weak at reducing the redundancy in scene graph data. This article introduces a novel lossless compression framework with adaptive predictors for the joint compression of objects and relations in scene graph data. The proposed framework comprises a unified prior extractor and specialized element predictors to adapt to different data elements. Furthermore, to exploit the context information within and between graph elements, Graph Context Convolution is proposed to support different graph context modeling schemes for different graph elements. Finally, an overarching framework incorporates the learned distribution model to predict numerical data under complicated conditional constraints. Experiments conducted on labeled or generated scene graphs demonstrate the effectiveness of the proposed framework for scene graph lossless compression.

The rapid advancement of multimedia and imaging technologies has resulted in increasingly diverse visual and semantic data. A large range of applications such as remote-assisted driving requires the amalgamated storage and transmission of various visual and semantic data. However, existing works suffer from the limitation of insufficiently exploiting the redundancy between different types of data. In this article, we propose a unified framework to jointly compress a diverse spectrum of visual and semantic data, including images, point clouds, segmentation maps, object attributes, and relations. We develop a unifying process that embeds the representations of these data into a joint embedding graph according to their categories, which enables flexible handling of joint compression tasks for various visual and semantic data. To fully leverage the redundancy between different data types, we further introduce an embedding-based adaptive joint encoding process and a Semantic Adaptation Module to efficiently encode diverse data based on the learned embeddings in the joint embedding graph. Experiments on the Cityscapes, MSCOCO, and KITTI datasets demonstrate the superiority of our framework, highlighting promising steps toward scalable multimedia processing.

There has been a recent surge of interest in introducing transformers to 3D human pose estimation (HPE) due to their powerful capabilities in modeling long-term dependencies. However, existing transformer-based methods treat body joints as equally important inputs and ignore the prior knowledge of human skeleton topology in the self-attention mechanism. To tackle this issue, in this paper, we propose a Pose-Oriented Transformer (POT) with uncertainty guided refinement for 3D HPE. Specifically, we first develop novel pose-oriented self-attention mechanism and distance-related position embedding for POT to explicitly exploit the human skeleton topology. The pose-oriented self-attention mechanism explicitly models the topological interactions between body joints, whereas the distance-related position embedding encodes the distance of joints to the root joint to distinguish groups of joints with different difficulties in regression. Furthermore, we present an Uncertainty-Guided Refinement Network (UGRN) to refine pose predictions from POT, especially for the difficult joints, by considering the estimated uncertainty of each joint with uncertainty-guided sampling strategy and self-attention mechanism. Extensive experiments demonstrate that our method significantly outperforms the state-of-the-art methods with reduced model parameters on 3D HPE benchmarks such as Human3.6M and MPI-INF-3DHP.

Domain Generalization (DG) has achieved great success in generalizing knowledge from source domains to unseen target domains. However, current DG methods rely heavily on labeled source data, which are usually costly and unavailable. Since unlabeled data are far more accessible, we study a more practical unsupervised domain generalization (UDG) problem. Learning invariant visual representation from different views, i.e., contrastive learning, promises well semantic features for in-domain unsupervised learning. However, it fails in cross-domain scenarios. In this paper, we first delve into the failure of vanilla contrastive learning and point out that semantic connectivity is the key to UDG. Specifically, suppressing the intra-domain connectiv-ity and encouraging the intra-class connectivity help to learn the domain-invariant semantic information. Then, we propose a novel unsupervised domain generalization approach, namely Dual Nearest Neighbors contrastive learning with strong Augmentation (DN2A). Our DN2A leverages strong augmentations to suppress the intra-domain connectivity and proposes a novel dual nearest neighbors search strategy to find trustworthy cross domain neighbors along with in-domain neighbors to encourage the intra-class connectivity. Experimental results demonstrate that our DN2A outperforms the state-of-the-art by a large margin, e.g., 12.01% and 13.11 % accuracy gain with only 1% labels for linear evaluation on PACS and DomainNet, respectively.

Pretraining followed by fine-tuning has proven to be effective in visual recognition tasks. However, fine-tuning all parameters can be computationally expensive, particularly for large-scale models. To mitigate the computational and storage demands, recent research has explored Parameter-Efficient Fine-Tuning (PEFT), which focuses on tuning a minimal number of parameters for efficient adaptation. Existing methods, however, fail to analyze the impact of the additional parameters on the model, resulting in an unclear and suboptimal tuning process. In this paper, we introduce a novel and effective PEFT paradigm, named SNF (Shortcut adaptation via Normalization Flow), which utilizes normalizing flows to adjust the shortcut layers. We highlight that layers without Lipschitz constraints can lead to error propagation when adapting to downstream datasets. Since modifying the over-parameterized residual connections in these layers is expensive, we focus on adjusting the cheap yet crucial shortcuts. Moreover, learning new information with few parameters in PEFT can be challenging, and information loss can result in label information degradation. To address this issue, we propose an information-preserving normalizing flow. Experimental results demonstrate the effectiveness of SNF. Specifically, with only 0.036M parameters, SNF surpasses previous approaches on both the FGVC and VTAB-1k benchmarks using ViT/B-16 as the backbone. The code is available at https://github.com/Wang-Yaoming/SNF

End-to-end learned image compression (LIC) has become promising alternatives for lossy image compression. However, deployments of LIC models are limited in real-world applications due to excessive network parameters and high computational complexity. Existing LIC models realized with integer networks are significantly de-graded in rate-distortion (R-D) performance. In this paper, we propose a novel fully integerized LIC model that simultaneously achieves channel-wise weight and 8-bit activation quantization for alleviating the loss of R-D performance. For weight quantization, we develop an internal bit width increment (IBWI) via nonlinear logarithmic mapping to convert the stored low-precision integer weights to high bit-width weights for inference. Moreover, outlier channel splitting (OCS) [1] is employed to address large outliers of weight distribution by duplicating channels and constrain the integer weights within the scope of INT8. For activation quantization, we leverage activation equalization to balance the channel-wise distribution of activations. Experimental results demonstrate that the proposed method achieves a reduction of 75% storage cost with subtle performance loss compared to full-precision pre-trained models, and outperforms existing integer-only networks, as shown in Table 1.

Quantizing one single deep neural network into multiple compression rates (precisions) has been recently considered for flexible deployments in real-world scenarios. In this paper, we propose a novel scheme that achieves progressive bit-width allocation and joint training to simultaneously optimize mixed-precision quantized networks under multiple compression rates. Specifically, we develop a progressive bit-width allocation with switchable quantization step size to enable mixed-precision quantization based on analytic sensitivity of network layers under multiple compression rates. Furthermore, we achieve joint training for quantized networks under different compression rates via knowledge distillation to exploit their correlations based on the shared network structure. Experimental results show that the proposed scheme outperforms AdaBits [1] in various networks on CIFAR-10 and ImageNet.

Generalization ability of face anti-spoofing has been widely concerned in recent years. Existing domain generalization methods use adversarial learning or metric learning to extract invariant features across domains but are proved to be flawed from causal views. The learned domain-invariant features cannot generalize well to unseen real scenarios in both theoretical and practical senses. In this paper, we propose a novel method that learns Causal Representations for Face Anti-Spoofing (CRFAS). We first model the data generation process of face anti-spoofing via Structural Casual Model (SCM) and reveal that only the causal feature is capable of generalizing to unseen domains. On such basis, we extract causal features via back- door adjustment without prior assumptions rather than learn domain-invariant features as existing methods. Furthermore, we employ the supervised contrastive loss to generate more realistic counterfactual features for backdoor adjustment and improve the generalization ability of learned causal features. Extensive experiments on six cross-dataset testing scenarios demonstrate that CRFAS outperforms the state-of-the-art face anti-spoofing methods in terms of HTER and AUC in generalizing to unseen domains.

Generalizable face anti-spoofing (FAS) based on domain generalization (DG) has gained growing attention due to its robustness in real-world applications. However, these DG methods rely heavily on labeled source data, which are usually costly and hard to access. Comparably, unlabeled face data are far more accessible in various scenarios. In this paper, we propose the first Unsupervised Domain Generalization framework for Face Anti-Spoofing, namely UDG-FAS, which could exploit large amounts of easily accessible unlabeled data to learn generalizable features for enhancing the low-data regime of FAS. Yet without supervision signals, learning intrinsic live/spoof features from complicated facial information is challenging, which is even tougher in cross-domain scenarios due to domain shift. Existing unsupervised learning methods tend to learn identity-biased and domain-biased features as shortcuts, and fail to specify spoof cues. To this end, we propose a novel Split-Rotation-Merge module to build identity-agnostic local representations for mining intrinsic spoof cues and search the nearest neighbors in the same domain as positives for mitigating the identity bias. Moreover, we propose to search cross-domain neighbors with domain-specific normalization and merged local features to learn a domain-invariant feature space. To our best knowledge, this is the first attempt to learn generalized FAS features in a fully unsupervised way. Extensive experiments show that UDG-FAS significantly outperforms state-of-the-art methods on six diverse practical protocols.

Point cloud upsampling is a challenging task to generate dense and uniform point clouds from sparse and irregular inputs. Existing deep learning based methods usually leverage graph convolution networks (GCNs) to learn features for expansion and reconstruction but suffer from the over-smoothing issue by discarding the high-frequency components. In this paper, we design a novel dynamic graph scattering transform (GST) to achieve multi-scale feature extraction for point cloud upsampling. The dynamic GST can accommodate to the K-nearest neighbor (KNN)-based feature graph dynamically constructed for point clouds to achieve multi-resolution band-pass representations. Consequently, a feature extractor with dense connection is developed based on the dynamic GST such that low-frequency and high-frequency components can be sufficiently combined for subsequent feature expansion and reconstruction. To our best knowledge, this paper is the first to introduce graph scattering transform for representation learning with dynamic feature graphs in point cloud upsampling. Experimental results on the PU1K dataset show that the proposed network based on dynamic GST achieve promising performance under various metrics especially for sparse point clouds.

Recent 2D-to-3D human pose estimation (HPE) utilizes temporal consistency across sequences to alleviate the depth ambiguity problem but ignore the action related prior knowledge hidden in the pose sequence. In this paper, we propose a plug-and-play module named Action Prompt Module (APM) that effectively mines different kinds of action clues for 3D HPE. The highlight is that, the mining scheme of APM can be widely adapted to different frameworks and bring consistent benefits. Specifically, we first present a novel Action-related Text Prompt module (ATP) that directly embeds action labels and transfers the rich language information in the label to the pose sequence. Besides, we further introduce Action-specific Pose Prompt module (APP) to mine the position-aware pose pattern of each action, and exploit the correlation between the mined patterns and input pose sequence for further pose refinement. Experiments show that APM can improve the performance of most video-based 2D-to-3D HPE frameworks by a large margin.

In personalized federated learning (PFL), multiple clients train customized models to fulfill their personal objectives, which, however, are prone to overfitting to local data due to the heterogeneity and scarcity of local data. To address this, we propose from the information-theoretic perspective a personalized federated learning framework based on the common representation learned across clients, named FedCR. Specifically, we introduce to the local client update a regularizer that aims at minimizing the discrepancy between local and global conditional mutual information (CMI), such that clients are encouraged to learn and exploit the common representation. Upon this, each client learns individually a customized predictor (head), while the extractor (body) remains to be aggregated by the server. Our CMI regularizer leads to a theoretically sound alignment between the local and global stochastic feature distributions in terms of their Kullback-Leibler (KL) divergence. More importantly, by modeling the global joint feature distribution as a product of multiple local feature distributions, clients can efficiently extract diverse information from the global data but without need of the raw data from other clients. We further show that noise injection via feature alignment and ensemble of local predictors in FedCR would help enhance its generalization capability. Experiments on benchmark datasets demonstrate a consistent performance gain and better generalization behavior of FedCR.

Learning a rotation-invariant (RI) representation is of significant importance for real-world point cloud segmentation that is perturbed by arbitrary rotations. Recent principal component analysis (PCA)-based methods provide an effective alternative to align point clouds and produce the RI representation under the preservation of global information. However, conventional PCA with 3-D coordinates cannot fully represent high-dimensional geometric structures like surfaces and curves and cannot uniformly align these structures for learning RI representation. In this paper, we propose a novel rotation-invariant method for point cloud segmentation, which leverages kernel PCA (KPCA) for aligning point clouds in a projected high-dimensional space via non-linear mapping and develops a Geometry-based Weighted Convolution (GWConv) to distinguish part boundaries during segmentation. Specifically, the KPCA produces a RI representation with polynomial kernels for effectively representing complicated geometric structures in point clouds. Moreover, the GWConv incorporates geometric structures into convolution and enhances neighboring points with similar geometry for fine-grained segmentation based on the RI representation. Experimental results demonstrate that the proposed method can achieve competitive performance with the state-of-the-arts and outperforms existing PCA-based methods in part segmentation on ShapeNet. Furthermore, it achieves evident performance gains on complicated 3-D shapes such as Earphone and Car and facilitates segmentation around the part boundaries.

Parameter-Efficient Tuning (PET) has emerged as a leading advancement in both Natural Language Processing and Computer Vision, enabling efficient accommodation of downstream tasks without costly fine-tuning. However, most existing PET approaches are limited to uni-modal tuning, even for vision-language models like CLIP. We investigate this limitation and demonstrate that simultaneous tuning of the two modalities in such models leads to multi-modal forgetting and catastrophic performance degradation, particularly when generalizing to new classes. To address this issue, we propose a novel PET approach called VioLET (Vision Language Efficient Tuning) that utilizes collaborative multi-modal gradients to unlock the full potential of both modalities. Specifically, we incorporate an additional visual encoder without learnable parameters and use these two visual encoders to compute the gradients of the context parameters separately. When conflicts arise, we replace the original gradient with an orthogonal gradient. Extensive experiments are conducted on few-shot recognition and unseen class generalization tasks using ResNet-50 or ViT/B-16 as the backbone. VioLET consistently outperforms several state-of-the-art methods on 11 datasets, showcasing its superiority over existing PET approaches. The code is available at https://github.com/Wang-Yaoming/VioLET.

Meta-reinforcement learning (Meta-RL), though enabling a fast adaptation to learn new skills by exploiting the common structure shared among different tasks, suffers performance degradation in the sparse-reward setting. Current hindsight-based sample transfer approaches can alleviate this issue by transferring relabeled trajectories from other tasks to a new task so as to provide informative experience for the target reward function, but are unfortunately constrained with the unrealistic assumption that tasks differ only in reward functions. In this paper, we propose a doubly robust augmented transfer (DRaT) approach, aiming at addressing the more general sparse reward meta-RL scenario with both dynamics mismatches and varying reward functions across tasks. Specifically, we design a doubly robust augmented estimator for efficient value-function evaluation, which tackles dynamics mismatches with the optimal importance weight of transition distributions achieved by minimizing the theoretically derived upper bound of mean squared error (MSE) between the estimated values of transferred samples and their true values in the target task. Due to its intractability, we then propose an interval-based approximation to this optimal importance weight, which is guaranteed to cover the optimum with a constrained and sample-independent upper bound on the MSE approximation error. Based on our theoretical findings, we finally develop a DRaT algorithm for transferring informative samples across tasks during the training of meta-RL. We implement DRaT on an off-policy meta-RL baseline, and empirically show that it significantly outperforms other hindsight-based approaches on various sparse-reward MuJoCo locomotion tasks with varying dynamics and reward functions.

Vision transformers (ViTs) have emerged as a prevalent architecture for vision tasks owing to their impressive performance. However, their complexity dramatically increases when handling long token sequences, particularly for dense prediction tasks that require high-resolution input. Notably, dense prediction tasks, such as semantic segmentation or object detection, emphasize more on the contours or shapes of objects, while the texture inside objects is less informative. Motivated by this observation, we propose to apply adaptive resolution for different regions in the image according to their importance. Specifically, at the intermediate layer of the ViT, we select anchors from the token sequence using the proposed spatial-aware density-based clustering algorithm. Tokens that are adjacent to anchors are merged to form low-resolution regions, while others are preserved independently as high-resolution. This strategy could significantly reduce the number of tokens, and the following layers only handle the reduced token sequence for acceleration. At the output end, the resolution of the feature map is recovered by unfolding merged tokens for task prediction. Consequently, we can considerably accelerate ViTs for dense prediction tasks. The proposed method is evaluated across three different datasets and demonstrates promising performance. For instance, "Segmenter ViT-L" can be accelerated by 48\% FPS without fine-tuning, while maintaining the performance. Moreover, our method can also be applied to accelerate fine-tuning. Experiments indicate that we can save 52\% training time while accelerating 2.46$\times$ FPS with only a 0.09\% performance drop.

Deep learning models have greatly improved the accuracy of point cloud classification. Nevertheless, existing deep learning models are vulnerable to data corruptions such as Gaussian noise and outliers which are inevitable in real-world point cloud collection. To address this challenge, we develop a novel point cloud classification model that is robust to data corruptions, with permutohedral lattice-based representation and density-improved hierarchical feature extraction. Specifically, raw noisy point clouds are firstly projected into a regular per-mutohedral lattice space to obtain the quantized representations. Subsequently, we propose to leverage the local density to improve the farthest point sampling (FPS) and spectral graph convolution for robust hierarchical feature extraction, inspired by that the local point density implicitly reveals the reliability and semantic information of each point. The local density can be efficiently calculated from the permutohedral lattice-based representation. Extensive experiments on the benchmark dataset (i.e., ModelNet-C) verify the robustness of the proposed model. In comparison to the state-of-the-art methods, the proposed model achieves superior performance on the corrupted dataset while maintaining competitive classification accuracy on the clean dataset.

Entropy modeling plays an important role in estimating the rates of latent representations and optimizing the rate-distortion performance for learned image compression. Autoregression modules are demonstrated to eliminate spatial/channel-wise redundancy of latent representations in fixed-rate learned image compression. However, it cannot be efficiently achieved in progressive coding due to the high computational complexity raised by element-wise probability prediction. In this paper, we propose a learned progressive image compression method that enables spatial autoregression for entropy modeling. Specifically, we develop a novel codeword alignment scheme to prevent coding redundancy and achieve efficient autoregression of latent representations in different quality layers. Consequently, conditional probability estimation for the latent prediction can be achieved based on spatial autoregression in a layer-wise manner. We further extend the proposed method by dead-zone quantizers to obtain promoted rate-distortion performance. The proposed method is a successful attempt to enable spatial autoregression in learned progressive coding and further bridge the performance gap with fixed-rate models. Experimental results show that it outperforms traditional methods such as JPEG and BPG, as well as recent fine-grained learned progressive coding models DPICT and PLONQ in terms of rate-distortion performance.

Batch normalization (BN) is a fundamental unit in modern deep neural networks. However, BN and its variants focus on normalization statistics but neglect the recovery step that uses linear transformation to improve the capacity of fitting complex data distributions. In this paper, we demonstrate that the recovery step can be improved by aggregating the neighborhood of each neuron rather than just considering a single neuron. Specifically, we propose a simple yet effective method named batch normalization with enhanced linear transformation (BNET) to embed spatial contextual information and improve representation ability. BNET can be easily implemented using the depth-wise convolution and seamlessly transplanted into existing architectures with BN. To our best knowledge, BNET is the first attempt to enhance the recovery step for BN. Furthermore, BN is interpreted as a special case of BNET from both spatial and spectral views. Experimental results demonstrate that BNET achieves consistent performance gains based on various backbones in a wide range of visual tasks. Moreover, BNET can accelerate the convergence of network training and enhance spatial information by assigning important neurons with large weights accordingly.

Self-supervised learning based on instance discrimination has shown remarkable progress. In particular, contrastive learning, which regards each image as well as its augmentations as an individual class and tries to distinguish them from all other images, has been verified effective for representation learning. However, conventional contrastive learning does not model the relation between semantically similar samples explicitly. In this paper, we propose a general module that considers the semantic similarity among images. This is achieved by expanding the views generated by a single image to Cross-Samples and Multi-Levels, and modeling the invariance to semantically similar images in a hierarchical way. Specifically, the cross-samples are generated by a data mixing operation, which is constrained within samples that are semantically similar, while the multi-level samples are expanded at the intermediate layers of a network. In this way, the contrastive loss is extended to allow for multiple positives per anchor, and explicitly pulling semantically similar images together at different layers of the network. Our method, termed as CSML, has the ability to integrate multi-level representations across samples in a robust way. CSML is applicable to current contrastive based methods and consistently improves the performance. Notably, using MoCo v2 as an instantiation, CSML achieves 76.6% top-1 accuracy with linear evaluation using ResNet-50 as backbone, 66.7% and 75.1% top-1 accuracy with only 1% and 10% labels, respectively. All these numbers set the new state-of-the-art. The code is available at https://github.com/haohang96/CSML.

It is promising to solve linear inverse problems by unfolding iterative algorithms (e.g., iterative shrinkage thresholding algorithm (ISTA)) as deep neural networks (DNNs) with learnable parameters. However, existing ISTA-based unfolded algorithms restrict the network architectures for iterative updates with the partial weight coupling structure to guarantee convergence. In this paper, we propose hybrid ISTA to unfold ISTA with both pre-computed and learned parameters by incorporating free-form DNNs (i.e., DNNs with arbitrary feasible and reasonable network architectures), while ensuring theoretical convergence. We first develop HCISTA to improve the efficiency and flexibility of classical ISTA (with pre-computed parameters) without compromising the convergence rate in theory. Furthermore, the DNN-based hybrid algorithm is generalized to popular variants of learned ISTA, dubbed HLISTA, to enable a free architecture of learned parameters with a guarantee of linear convergence. To our best knowledge, this paper is the first to provide a convergence-provable framework that enables free-form DNNs in ISTA-based unfolded algorithms. This framework is general to endow arbitrary DNNs for solving linear inverse problems with convergence guarantees. Extensive experiments demonstrate that hybrid ISTA can reduce the reconstruction error with an improved convergence rate in the tasks of sparse recovery and compressive sensing.

Progressive coding is essential to the practical deployment of learned image compression over heterogeneous networks and clients. Existing methods for learned progressive image compression require complex and empirical design to achieve near-optimal rate-distortion performance over a wide range of bit-rates. However, these methods are limited by the implicit learned mechanism based on neural networks and introduction of uniform quantizers. In this paper, we propose generalized learned progressive image compression with analytic rate-distortion optimization using dead-zone quantizers on the latent representation. Specifically, we reveal that dead-zone quantizers, as a general case of uniform quantizers, are equivalent to uniform quantizers in fixed-rate nonlinear transform coding and can prevent extra redundancy in embedded quantization for progressive coding. Consequently, we propose rate-distortion optimized learned progressive coding by approximating the optimal quantizer in the source spaces using dead-zone quantizers in an analytic manner on the Laplacian source. To our best knowledge, this paper is the first to achieve general learned progressive coding from the perspective of optimal quantizers. The proposed method achieves theoretically sound and practically efficient embedded quantization and learned progressive coding of latent representations with improved rate-distortion performance. It can also enable embedded quantization with diverse assignments of truncation points and support flexible configuration of quality layers of varying numbers and at varying target bit-rates. Furthermore, we successfully incorporate the proposed method into existing pre-trained fixed-rate models to realize progressive learned image compression without re-training. Experimental results demonstrate that the proposed method achieves state-of-the-art rate-distortion performance in learned progressive image compression compared with traditional codecs and recent learned methods.

Endobronchial ultrasound (EBUS) elastography videos have shown great potential to supplement intrathoracic lymph node diagnosis. However, it is laborious and subjective for the specialists to select the representative frames from the tedious videos and make a diagnosis, and there lacks a framework for automatic representative frame selection and diagnosis. To this end, we propose a novel deep learning framework that achieves reliable diagnosis by explicitly selecting sparse representative frames and guaranteeing the invariance of diagnostic results to the permutations of video frames. Specifically, we develop a differentiable sparse graph attention mechanism that jointly considers frame-level features and the interactions across frames to select sparse representative frames and exclude disturbed frames. Furthermore, instead of adopting deep learning-based frame-level features, we introduce the normalized color histogram that considers the domain knowledge of EBUS elastography images and achieves superior performance. To our best knowledge, the proposed framework is the first to simultaneously achieve automatic representative frame selection and diagnosis with EBUS elastography videos. Experimental results demonstrate that it achieves an average accuracy of 81.29% and area under the receiver operating characteristic curve (AUC) of 0.8749 on the collected dataset of 727 EBUS elastography videos, which is comparable to the performance of the expert-based clinical methods based on manually-selected representative frames.

With increasing popularity of virtual reality and augmented reality, application of point clouds is in critical demand as it enables users to freely navigate in an immersive scene with six degrees of freedom. However, point clouds usually comprise large amounts of data, and are thus difficult to stream in bandwidth-constrained networks. It is therefore important, yet challenging, to efficiently stream the resource-intensive point clouds, such that the user’s quality of experience (QoE) is guaranteed on a high-level but with a low bandwidth consumption. To this end, we propose a QoE-driven adaptive streaming approach for the tile-based point cloud transmission, to maximize the user’s QoE while reducing the transmission redundancy. By exploiting the perspective projection, we specifically model the QoE of a 3D tile as a function of the bitrate of its representation, user’s view frustum and spatial position, occlusion between tiles, and the resolution of rendering device. Based on this QoE model, we then formulate the QoE-optimized rate adaptation problem as a multiple-choice knapsack problem, which allocates bitrates for different tiles under a given transmission capacity. It is equivalently converted to a submodular function maximization problem subject to knapsack constraints, and solved by a practical greedy-based algorithm with a theoretical worst-case performance guarantee. The proposed algorithm is able to achieve a near-optimal performance, but with a very low computational complexity. Experimental results further demonstrate superiority of the proposed rate adaptation algorithm over existing schemes, in terms of both user’s visual quality and transmission efficiency.

Light field (LF) data are widely used in the immersive representations of the 3D world. To record the light rays along with different directions, an LF requires much larger storage space and transmission bandwidth than a conventional 2D image with similar spatial dimension. In this paper, we propose a novel framework for light field image compression that leverages graph learning and dictionary learning to remove structural redundancies between different views. Specifically, to significantly reduce the bit-rates, only a few key views are sampled and encoded, whereas the remaining non-key views are reconstructed via the graph adjacency matrix learned from the angular patch. Furthermore, dictionary-guided sparse coding is developed to compress the graph adjacency matrices and reduce the coding overheads. To our best knowledge, this paper is the first to achieve compact representation of cross-view structural information via adaptive learning on graphs. Experimental results demonstrate that the proposed framework achieves better performance than the standardized HEVC-based codec.

Appearance-based Gaze Estimation leverages deep neural networks to regress the gaze direction from monocular images and achieve impressive performance. However, its success depends on expensive and cumbersome annotation capture. When lacking precise annotation, the large domain gap hinders the performance of trained models on new domains. In this paper, we propose a novel gaze adaptation approach, namely Contrastive Regression Gaze Adaptation (CRGA), for generalizing gaze estimation on the target domain in an unsupervised manner. CRGA leverages the Contrastive Domain Generalization (CDG) module to learn the stable representation from the source domain and leverages the Contrastive Self-training Adaptation (CSA) module to learn from the pseudo labels on the target domain. The core of both CDG and CSA is the Contrastive Regression (CR) loss, a novel contrastive loss for regression by pulling features with closer gaze directions closer together while pushing features with farther gaze directions farther apart. Experimentally, we choose ETH-XGAZE and Gaze-360 as the source domain and test the domain generalization and adaptation performance on MPIIGAZE, RT-GENE, Gaze-Capture, EyeDiap respectively. The results demonstrate that our CRGA achieves remarkable performance improvement compared with the baseline models and also outperforms the state-of-the-art domain adaptation approaches on gaze adaptation tasks.

Existing entropy models in learned image compression are cumbersome to generate fixed mean and variance for estimating Gaussian distributions for latent representation. In this paper, we propose a novel entropy model based on Gaussian process regression (GPR) that flexibly predicts the mean of Gaussians with posterior distributions characterized by covari-ance functions spanned in the high-dimensional feature space. Furthermore, we develop the rate-distortion optimization based on the proposed entropy model by approximating the bitrates with the evidence lower bound (ELBO) derived via variational inference for GPR. The proposed model can be seamlessly integrated into existing end-to-end optimized frame-works by substituting the masked convolution based autoregressive models. Experimental results demonstrate that the proposed model outperforms conventional image compression methods such as JPEG2000 and BPG, as well as recent learning based methods on the Kodak dataset in terms of rate-distortion performance.

Existing methods for cross-component prediction focus on the chroma intra pre-diction but neglect the chroma prediction across multiple frames. In this paper, we propose a novel compression framework that leverages chroma frame sampling and inter-frame chroma prediction to address this problem. Specifically, chroma frame sampling discards the chroma components in inter-predicted frames to further reduce the bit consumption, whereas inter-frame chroma prediction recovers the discarded chroma components with optical flow estimation and post-training optimization for a guaranteed fidelity. To our best knowledge, this paper is the first attempt to achieve deep learning-based inter prediction of chroma components. To accommodate the proposed framework, we redesign the HEVC codec to enable hybrid coding of luma and chroma components. Experimental results show that the proposed framework achieves up to 0.76% BD-rate reduction when compared to standard HEVC.

3-D point clouds rendering solid representations of scenes or objects often carry a tremendous amount of points, compulsorily requesting high-efficiency compression for storage and transmission. In this paper, we propose a novel$p$-Laplacian embedding graph dictionary learning algorithm for 3-D point cloud attribute compression. The proposed method integrates the underlying graph topology to the learned graph dictionary capitalizing on$p$-Laplacian eigenfunctions and leads to parsimonious representations of 3-D point clouds. We further devise alternating optimization with the help of ADMM to efficiently solve the resulting non-convex minimization problem.

We propose a structured pruning method to achieve a light-weighted decoder of learned image compression to accommodate various terminals. The structured pruning method identifies the effectiveness of each channel of decoder via gradient ascent and gradient descent while maintaining the encoder and entropy model. To our best knowledge, this paper is the first attempt to design a structured pruning method for universal pretrained learned image compression. Experimental results demonstrate that the proposed method can reduce about 40% parameters and save 25% inference time at the cost of 0.23 dB BD-PSNR and 4.33% BD-rate change.

Generalizable face anti-spoofing (FAS) has drawn growing attention due to its robustness to unseen real scenarios. Existing domain generalization methods leverage adversarial learning or meta-learning to mitigate the domain bias and improve generalizability. However, these methods are heuristic and suffer from complicated min-max problems or cumbersome meta-updates. In this paper, we propose a simple yet effective Causal Intervention method for generalizable Face Anti-Spoofing, namely CIFAS. Firstly, we figure out the generalizability is undermined by a domain-aware confounder based on the structural causal model. Instantiating the confounder as the domain-specific factor, a domain embedding module is employed with Dirichlet mixup to obtain representative domain features. Consequently, we propose a novel backdoor adjustment model for causal intervention to capture the true causality and learn a robust FAS model. Our CIFAS is the first attempt to introduce causal learning into FAS. Extensive experiments on seven cross-dataset tests demonstrate that CIFAS outperforms the state-of-the-art methods.

Network quantization can facilitate the practical deployment of U-Net in various medical applications, especially medical image segmentation. However, existing quantization methods for U-Net do not sufficiently exploit the skip connection between its encoder and decoder structures. In this paper, we propose a novel mixed-precision quantization scheme for U-Net that leverages split convolution and skip-supervised quantization aware training to address this problem. Specifically, split convolution decouples the feature maps obtained by skip connection and up-sampling to enable fine-grained bitwidth allocation for mixed-precision quantization. Furthermore, we introduce supervision on skip connection for quantization aware training to compensate gradients in back propagation and enhance the skip features in finetuning. To our best knowledge, this is the first attempt to design mixed-precision quantization specifically for medical image segmentation with U-Net. Experimental results on four medical image datasets demonstrate that the proposed scheme outperforms existing U-Net quantization strategy and recent general-purpose mixed-precision quantizaton methods in medical image segementation.

Recent video colorization methods optimize correspondence estimation and information propagation in an end-to-end manner. However, they usually suffer from loss of fidelity due to the inaccurate inference of correspondence measurement. In this paper, we propose a post-training optimization (PTO) strategy to refine correspondence measurement in the end-to-end optimized framework. The proposed PTO strategy introduces a pseudo loss function to well approximate the target loss and guide the direction of updates. We further develop a video colorization method that incorporates PTO and optical flow to guarantee high-fidelity colorized frames in theory. Experimental results demonstrate that the proposed method achieves state-of-the-art PSNR performance in video colorization on the DAVIS dataset and common test sequences for video coding. Furthermore, the proposed method can be employed into video compression and achieves competitive rate-distortion performance with the recent High Efficiency Video Coding (HEVC) standard.

Differentiable architecture search (DARTS) has achieved success in searching powerful network architectures but suffers from unstable search, high search cost from repetitive attempts and unawareness of computational cost. In this paper, we propose a novel approach, namely Efficient and Stable Differentiable Architecture Search (ES-DARTS), that leverages decoupled search strategy and variational proxy pruning to achieve efficient search for efficient neural networks. Specifically, the decoupled search strategy stabilizes the search via bridging the gap between search and evaluation, and achieves acceleration with decoupled optimization, while the variational proxy pruning introduces structural pruning into DARTS to accommodate varying constraints on computational cost. ES-DARTS can be a plug-and-play module for DARTS-based approaches to achieve improved performance with reduced search cost and FLOPS. Experimental results demonstrate that ES-DARTS reduces top-1 error rate to 2.71% with $10 \times$ acceleration of DARTS on CIFAR10. ES-DARTS finds an architecture of 2.80% top-1 error rate with only 2.27 MB parameters on CIFAR-10 and of 29.4% top-1 error rate with only 322M FLOPS when transferred to ImageNet.

In this paper, we present a meta reinforcement learning (Meta-RL)-based neural adaptive bitrate streaming (ABR) algorithm that is able to rapidly adapt its control policy to the changing network throughput dynamics. Specifically, to allow rapid adaptation, we discuss the necessity of detaching the inference of throughput dynamics with the universal control mechanism that is in essence shared by all potential throughput dynamics for neural ABR algorithms. To meta-learn the ABR policy, we then build up a model-free system framework, composed of a probabilistic latent encoder that infers the underlying dynamics from the recent throughput context, and a policy network that is conditioned on latent variable and learns to quickly adapt to new environments. Additionally, to address the difficulties caused by training the policy on mixed dynamics, on-policy RL (or imitation learning) algorithms are suggested for policy training, with a mutual information-based regularization to make the latent variable more informative about the policy. Finally, we implement our algorithm's meta-training and meta-adaptation procedures under a variety of throughput dynamics. Empirical evaluations on different QoE metrics and multiple datasets containing real-world network traces demonstrate that our algorithm outperforms state-of-the-art ABR algorithms, in terms of the performance on the average chunk QoE, consistency and fast adaptation across a wide range of throughput patterns.

Entropy coding is fundamental for reducing the coding redundancy in image compression. However, existing entropy models for learned image compression are restricted by independent or autoregressive modeling based on presumed distribution generated from the family of Gaussian functions. In this paper, we propose a novel tensor network-based entropy model that can explicitly infer the joint distribution of image representation for learned image compression. We utilize a tree tensor network (TTN) to enable exact computation for probabilities of image representation. Specifically, we produce efficient tensor representation for entropy modeling based on bit planes slicing with gray code. Furthermore, we leverage tensor contraction to accurately calculate the partition function and jointly predict the entire bit plane for entropy coding. To our best knowledge, this paper is the first to directly infer the joint distribution of image representation in learned image compression without any presumed condition. Experimental results demonstrate that the proposed method is competitive with the state-of-the-art in both lossless and lossy image compression.

360-degree videos pose a challenge to their transmission due to the ultra-high resolution. To achieve efficient streaming, current tile-based schemes divide the 2D projection of spherical frames into small rectangular tiles, which are treated as individual sources for encoding and then delivered according to user's actual viewport. However, the trade-off between transmission and encoding efficiency, as controlled by encoding parameters such as tile sizes and quantization steps, has not been fully studied yet. In this paper, we propose a tile-based encoding scheme to strike this trade-off, by selecting the optimal encoding parameters that minimize the distortion of video displayed for user's viewport under a given transmission capacity. Specifically, we quantify the transmission efficiency by estimating the average number of pixels required for transmission to cover the user's viewport, while a CNN-based predictor is trained to predict the encoding efficiency based on video content characteristics, expressed as number of bits needed for encoding one pixel. We then formulate a rate-distortion optimization problem for tile-based encoding, which is solved efficiently by decreasing the feasible set of encoding parameters. Simulation results demonstrate that the proposed optimal tile-based encoding outperforms both fixed tiling and state-of-the-art adaptive tiling schemes in terms of rate-distortion behavior.

Semi-supervised learning, which assigns pseudo labels with models trained using limited labeled data, has been widely used in object detection to reduce the labeling cost. However, the provided pseudo annotations inevitably suffer noise since the initial model is not perfect. To address this issue, this paper introduces contrastive learning into semi-supervised object detection, and we claim that contrastive loss, which inherently relies on data augmentations, is much more robust than traditional softmax regression for noisy labels. To take full advantage of it in the detection task, we incorporate labels prior to contrastive loss and leverage plenty of region proposals to enhance diversity, which is crucial for contrastive learning. In this way, the model is optimized to make the region-level features with the same class be translation and scale invariant. Furthermore, we redesign the negative memory bank in contrastive learning to make the training more efficient. As far as we know, we are the first attempt that introduces contrastive learning in semi-supervised object detection. Experimental results on detection benchmarks demonstrate the superiority of our method. Notably, our method achieves 79.9% accuracy on VOC, which is 6.2% better than the supervised baseline and 0.7% improvement compared with the state-of-the-art method.

Tile-based rate adaption can improve the quality of experience (QoE) for adaptive 360-degree video streaming under constrained network conditions, which, however, is a challenging problem due to the requirements of accurate prediction for users’ viewports and optimal bitrate allocation for tiles. In this paper, we propose a strategy that deploys reinforcement learning-based Rate Adaptation with adaptive Prediction and Tiling for 360-degree video streaming, named RAPT360, to address these challenges. Specifically, to improve the accuracy of the state-of-the-art viewport prediction approaches, we fit the time-varying Laplace distribution-based probability density function of the prediction error for different prediction lengths. On the basis of that, we develop a viewport identification method to determine the viewport area of a user depending on the buffer occupancy, where the obtained viewport can cover the real viewport with any given probability confidence level. We then propose a viewport-aware adaptive tiling scheme to improve the bandwidth efficiency, where three types of tile granularities are allocated according to the shape and position of the 2-D projection of that viewport. By establishing an adaptive streaming model and QoE metric specific to 360-degree videos, we finally formulate the rate adaptation problem for tile-based 360-degree video streaming as a non-linear discrete optimization problem that targets at maximizing the long-term user QoE under a bandwidth-constrained network. To efficiently solve this problem, we model the rate adaptation logic as a Markov decision process (MDP) and employ the deep reinforcement learning (DRL)-based algorithm to dynamically learn the optimal bitrate allocation of tiles. Extensive experimental results show that RAPT360 achieves a performance gain of at least 1.47 dB on average chunk QoE, including a video quality improvement of at least 1.33 dB, in comparison to the existing strategies for tile-based adaptive 360-degree video streaming.

The Cox proportional hazards model is a popular semi-parametric model for survival analysis. In this paper, we aim at developing a federated algorithm for the Cox proportional hazards model over vertically partitioned data (i.e., data from the same patient are stored at different institutions). We propose a novel algorithm, namely VERTICOX, to obtain the global model parameters in a distributed fashion based on the Alternating Direction Method of Multipliers (ADMM) framework. The proposed model computes intermediary statistics and exchanges them to calculate the global model without collecting individual patient-level data. We demonstrate that our algorithm achieves equivalent accuracy for the estimation of model parameters and statistics to that of its centralized realization. The proposed algorithm converges linearly under the ADMM framework. Its computational complexity and communication costs are polynomially and linearly associated with the number of subjects, respectively. Experimental results show that VERTICOX can achieve accurate model parameter estimation to support federated survival analysis over vertically distributed data by saving bandwidth and avoiding exchange of information about individual patients. The source code for VERTICOX is available at: https://github.com/daiwenrui/VERTICOX.

Model quantization is essential to deploy deep convolutional neural networks (DCNNs) on resource-constrained devices. In this article, we propose a general bitwidth assignment algorithm based on theoretical analysis for efficient layerwise weight and activation quantization of DCNNs. The proposed algorithm develops a prediction model to explicitly estimate the loss of classification accuracy led by weight quantization with a geometrical approach. Consequently, dynamic programming is adopted to achieve optimal bitwidth assignment on weights based on the estimated error. Furthermore, we optimize bitwidth assignment for activations by considering the signal-to-quantization-noise ratio (SQNR) between weight and activation quantization. The proposed algorithm is general to reveal the tradeoff between classification accuracy and model size for various network architectures. Extensive experiments demonstrate the efficacy of the proposed bitwidth assignment algorithm and the error rate prediction model. Furthermore, the proposed algorithm is shown to be well extended to object detection.

With the advent of data science, the analysis of network or graph data has become a very timely research problem. A variety of recent works have been proposed to generalize neural networks to graphs, either from a spectral graph theory or a spatial perspective. The majority of these works, however, focus on adapting the convolution operator to graph representation. At the same time, the pooling operator also plays an important role in distilling multiscale and hierarchical representations, but it has been mostly overlooked so far. In this article, we propose a parameter-free pooling operator, called iPool, that permits to retain the most informative features in arbitrary graphs. With the argument that informative nodes dominantly characterize graph signals, we propose a criterion to evaluate the amount of information of each node given its neighbors and theoretically demonstrate its relationship to neighborhood conditional entropy. This new criterion determines how nodes are selected and coarsened graphs are constructed in the pooling layer. The resulting hierarchical structure yields an effective isomorphism-invariant representation of networked data on arbitrary topologies. The proposed strategy achieves superior or competitive performance in graph classification on a collection of public graph benchmark data sets and superpixel-induced image graph data sets.

Cloud service has been emerging as a promising alternative to handle massive volumes of video sequences triggered by increasing demands on video service, especially surveillance and entertainment. Lossless compression of encoded video bitstreams can further eliminate the redundancies without altering the contents and facilitate the efficiency of cloud storage. In this paper, we propose a novel lossless compression scheme to further compress the video bitstreams generated by the state-of-the-art hybrid coding frameworks like H.264/AVC and HEVC. Different from transcoding, the proposed scheme develops frequency prediction and Macro Block (MB) merging to eliminate the redundancies remained in the intra-and inter-predicted frames with a strict guarantee of video fidelity. To our best knowledge, this paper is the first attempt to realize lossless compression of video bitstreams generated by advanced coding standards H.264/AVC and HEVC. Experimental results demonstrate that the proposed scheme can achieve compression gains of 17.38% and 10.85% on standard test sequences and surveillance videos, respectively.

Deep learning has been widely adopted in compression sensing (CS) to achieve superior reconstruction quality, but is restricted by the black-box architecture in network design and lack of interpretability. In this paper, we propose a novel deep network-based CS framework via unfolding the$\ell_{0}$-constrained convolutional sparse coding (CSC). The proposed method incorporates deep neural networks (DNNs) into the state-of-the-art optimization paradigm of alternating direction method of multipliers (ADMM). CS reconstruction can be improved with the compact and shift-invariant sparse representation under the merits of CSC and DNNs with learnable parameters. Convergence analysis demonstrates that the proposed method is guaranteed to converge, when the network parameters are reused for all the building blocks. Experimental results demonstrate that the proposed method is competitive or outperform the state-of-the-art DNN-based methods in CS of natural images and magnetic resonance imaging (MRI).

Application of point clouds is in critical demand, which, however, are composed of large amounts of data and difficult to stream in bandwidth-constrained networks. To address this, we propose a QoE-driven and tile-based adaptive streaming approach for point clouds, to reduce transmission redundancy and maximize user’s QoE. Specifically, by utilizing the perspective projection, we model the QoE of a 3D tile as a function of the bitrate of its representation, user’s view frustum and spatial position, occlusion between tiles, and the resolution of rendering device. We then formulate the QoE-optimized rate adaptation problem as a multiple-choice knapsack problem that allocates bitrates for different tiles under a given transmission capacity. We equivalently convert it as a submodular function maximization problem subject to knapsack constraints, and develop a practical greedy algorithm with a theoretical performance guarantee. Experimental results further demonstrate superiority of the proposed rate adaptation algorithm over existing schemes, in terms of both user’s visual quality and transmission efficiency.

Existing differentiable neural architecture search approaches simply assume the architectural distribution on each edge is independent of each other, which conflicts with the intrinsic properties of architecture. In this paper, we view the architectural distribution as the latent representation of specific data points. Then we propose Variational Information Maximization Neural Architecture Search (VIM-NAS) to leverage a simple yet effective convolutional neural network to model the latent representation, and optimize for a tractable variational lower bound to the mutual information between the data points and the latent representations. VIM-NAS automatically learns a nearly one-hot distribution from a continuous distribution with extremely fast convergence speed, e.g., converging with one epoch. Experimental results demonstrate VIM-NAS achieves state-of-the-art performance on various search spaces, including DARTS search space, NAS-Bench-1shot1, NAS-Bench-201, and simplified search spaces S1-S4. Specifically, VIM-NAS achieves a top-1 error rate of 2.45% and 15.80% within 10 minutes on CIFAR-10 and CIFAR-100, respectively, and a top-1 error rate of 24.0% when transferred to ImageNet.

Learned image compression has achieved improved rate-distortion performance with end-to-end optimized framework based on deep neural networks. However, context-based entropy modeling for learned image compression cannot simultaneously achieve enhanced efficiency and sufficiently exploiting the channel-wise correlations. In this paper, we propose a novel framework for learned image compression with channel-wise grouped context modeling. The proposed framework presents channel-wise grouping to explicitly exploit the channel-wise correlations and develop a grouped 3-D context model to achieve efficient entropy coding with a guarantee of rate-distortion performance. The proposed framework achieves competitive performance with a significantly reduced decoding complexity in comparison to 3-D context models.

State-of-the-art deep reinforcement learning (DRL) algorithms tend to overfit due to the model discrepancy between source and target environments. Though applying domain randomization during training can improve the average performance by randomly generating a sufficient diversity of environments in simulator, the worst-case environment is still neglected without any performance guarantee. Since the average and worst-case performance are both important for generalization in RL, in this paper, we propose a policy optimization approach for concurrently improving the policy’s performance in the average and worst-case environment. We theoretically derive a lower bound for the worst-case performance of a given policy by relating it to the expected performance. Guided by this lower bound, we formulate an optimization problem to jointly optimize the policy and sampling distribution, and prove that by iteratively solving it the worst-case performance is monotonically improved. We then develop a practical algorithm, named monotonic robust policy optimization (MRPO). Experimental evaluations in several robot control tasks demonstrate that MRPO can generally improve both the average and worst-case performance in the source environments for training, and facilitate in all cases the learned policy with a better generalization capability in some unseen testing environments.

In this paper, we propose BayesMPC, an uncertainty-aware robust adaptive bitrate (ABR) algorithm on the basis of Bayesian neural network (BNN) and model predictive control (MPC). Specifically, to improve the capacity of learning transition probability of the network throughput, we adopt a BNN-based predictor that is able to predict the statistical distribution of future throughput from the past throughput by not only considering the aleatoric uncertainty (e.g., noise), but also capturing the epistemic uncertainty incurred by lack of adequate training samples. We further show that by using the negative log-likelihood loss function to train this BNN-based throughput predictor, the generalization error can be minimized with the guarantee of PAC-Bayesian theorem. Rather than a point estimate, the learnt uncertainty can contribute to a confidence region for the future throughput, the lower bound of which then leads to an uncertainty-aware robust MPC strategy to maximize the worst-case user quality-of-experience (QoE) w.r.t. this confidence region. Finally, experimental results on three real-world network trace datasets validate the efficiency of both the proposed BNN-based predictor and uncertainty-aware robust MPC strategy, and demonstrate the superior performance compared to other baselines, in terms of both the overall QoE performance and generalization across all ranges of heterogeneous network and user conditions.

Mixed-precision quantization with adaptive bitwidth allocation for neural network has achieved higher compression rate and accuracy in classification task. However, it has not been well explored for object detection networks. In this paper, we propose a novel mixed-precision quantization scheme with dynamical Hessian matrix for object detection networks. We iteratively select a layer with the lowest sensitivity based on the Hessian matrix and downgrade its precision to reach the required compression ratio. The L-BFGS algorithm is utilized for updating the Hessian matrix in each quantization iteration. Moreover, we specifically design the loss function for objection detection networks by jointly considering the quantization effects on classification and regression loss. Experimental results on RetinaNet and Faster R-CNN show that the proposed DHMQ achieves state-of-the-art performance for quantized object detec-tors.

Differentiable architecture search (DARTS) enables effective neural architecture search (NAS) using gradient descent, but suffers from high memory and computational costs. In this paper, we propose a novel approach, namely Partially-Connected DARTS (PC-DARTS), to achieve efficient and stable neural architecture search by reducing the channel and spatial redundancies of the super-network. In the channel level, partial channel connection is presented to randomly sample a small subset of channels for operation selection to accelerate the search process and suppress the over-fitting of the super-network. Side operation is introduced for bypassing (non-sampled) channels to guarantee the performance of searched architectures under extremely low sampling rates. In the spatial level, input features are down-sampled to eliminate spatial redundancy and enhance the efficiency of the mixed computation for operation selection. Furthermore, edge normalization is developed to maintain the consistency of edge selection based on channel sampling with the architectural parameters for edges. Theoretical analysis shows that partial channel connection and parameterized side operation are equivalent to regularizing the super-network on the weights and architectural parameters during bilevel optimization. Experimental results demonstrate that the proposed approach achieves higher search speed and training stability than DARTS. PC-DARTS obtains a top-1 error rate of 2.55 percent on CIFAR-10 with 0.07 GPU-days for architecture search, and a state-of-the-art top-1 error rate of 24.1 percent on ImageNet (under the mobile setting) within 2.8 GPU-days.

Existing 3-D point cloud attribute compression methods are rigid for complex signals defined in spatially irregular domain, as they solely depend on geometric information or adopt pre-defined analytic bases for transform. In this article, we propose a multi-scale structured dictionary learning algorithm for 3-D point cloud attribute compression. The proposed multi-scale dictionary leverages hierarchical sparsity within signals by adapting atoms in an arborescent structure. It supports hierarchical sparse transform for encoding voxelized point cloud attributes with progressive refinement of high-frequency details from coarser to finer scales. To represent spatial blocks with varying amounts of occupied voxels, a non-uniform binary weight matrix is incorporated in dictionary learning to characterize the dimensional irregularity of signals. To guarantee optimal approximation with improved efficiency, alternating optimizations including Alternating Direction Method of Multipliers andGauss-Seideliterations are adopted for the weighted mixed$\ell _{2}/\ell _{1}$-minimization. Furthermore, we develop a lossy attribute compression framework by integrating sparse transform based prediction with dedicated quantization and entropy coding routines. Experimental results demonstrate that the proposed framework outperforms the state-of-the-art transform-based coding methods and is competitive with the most recent MPEG PCC test models in point cloud attribute compression.

Convolutional neural networks (CNNs) designed for low-dimensional regular grids will unfortunately lead to non-optimal solutions for analyzing spherical images, due to their different geometrical properties from planar images. In this paper, we generalize the grid-based CNNs to a non-Euclidean space by taking into account the geometry of spherical surfaces and propose a Spherical Graph Convolutional Network (SGCN) to encode rotation equivariant representations. Specifically, we propose a spherical graph construction criterion showing that a graph needs to be regular by evenly covering the spherical surfaces in order to design a rotation equivariant graph convolutional layer. For the practical case where the perfectly regular graph does not exist, we design two quantitative measures to evaluate the degree of irregularity for a spherical graph. The Geodesic ICOsahedral Pixelation (GICOPix) is adopted to construct spherical graphs with the minimum degree of irregularity compared to the current popular pixelation schemes. In addition, we design a hierarchical pooling layer to keep the rotation-equivariance, followed by a transition layer to enforce the invariance to the rotations for spherical image classification. We evaluate the proposed graph convolutional layers with different pixelations schemes in terms of equivariance errors. We also assess the effectiveness of the proposed SGCN in fulfilling rotation-invariance by the invariance error of the transition layers and recognizing the spherical images and 3D objects.

Generative model has emerged as a disruptive alternative for lossy compression of natural images, but suffers from the low-fidelity reconstruction. In this paper, we propose a noise-to-compression variational antoencoder (NC-VAE) to achieve efficient rate-distortion optimization (RDO) for end-to-end optimized image compression with a guarantee of fidelity. The proposed NC-VAE improves rate-distortion performance by adaptively adjusting the distribution of latent variables with trainable noise perturbation. Consequently, high-efficiency RDO is developed based on the distribution of latent variables for simplified decoder. Furthermore, robust end-to-end learning is developed over the corrupted inputs to suppress the deformation and color drift in standard VAE based generative models. Experimental results show that NC-VAE outperforms the state-of-the-art lossy image coders and recent end-to-end optimized compression methods in low bit-rate region, i.e., below 0.2 bits per pixel (bpp).

Reversible residual network naturally extends the linear lifting scheme with no theoretic guarantee. In this paper, we propose a reversible autoencoder (Rev-AE) with this extended non-linear lifting scheme to improve image reconstruction. Nonlinear prediction and update operators are designed based on shallow convolutional neural networks to model multilayer non-linearities. Different from existing autoencoders, Rev-AE support efficient image reconstruction with parameters reusable for the symmetric encoder and decoder. Rev-AE forms a set of related frames to guarantee perfect reconstruction with the non-linear extension of classic lifting scheme. Lower and upper bounds are developed for the set of frames to relate with the singular values for each non-linear operator. Furthermore, we employ Rev-AE into lossy image compression to evaluate its effectiveness on image reconstruction. Experimental results show that Rev-AE achieves competitive performance in comparison to the state-of-the-art.

Existing deep learning based methods for 3D human pose estimation cannot be deployed on resource-constrained devices. In this paper, we propose a Tiny-HourglassNet for 3D human pose estimation that improves the efficiency of Stacked Hourglass Networks with a guarantee of estimation performance. To be concrete, we develop an efficient Tiny-Hourglass backbone with two different types of ShuffleNet V2 blocks. At the meantime, we present a Simple Estimation Head (SEH) to reduce the resource utility for depth prediction. Furthermore, we develop two feature enhancement modules, namely Feature Enhancement Module (FEM) and Intermediate Enhancement Module (IEM), to improve the feature representation ability of Tiny-HourglassNet without evident increment of resource usage and computational complexity. Experimental results demonstrate that Tiny-HourglassNet achieves equivalent accuracy with a reduction of 80% complexity (FLOPs) in comparison to the baselines.

The k-nearest neighbors (KNN) algorithm has been widely adopted to construct graph convolutional networks (GCNs) for point cloud segmentation. However, the ℓ2norm cannot discriminate the multi-dimensional structures within a point cloud. In this paper, we propose a novel structure-aware graph construction for point clouds that compensates the ℓ2norm with per-dimension differences of the signal. The proposed method dynamically calculates the similarity ratio to determine the dimension-based proximity of the pair of points. Consequently, it improves both the spatial and spectral GCNs with the capability of aggregating information from relevant neighbors for point cloud segmentation. As a model-agnostic method, it can be seamlessly embedded into arbitrary GCN architectures during the graph construction phase. Experimental results demonstrate that the proposed method can improve classification accuracy around the joint areas of objects.

Bayesian methods have improved the interpretability and stability of neural architecture search (NAS). In this paper, we propose a novel probabilistic approach, namely Semi-Implicit Variational Dropout one-shot Neural Architecture Search (SI-VDNAS), that leverages semi-implicit variational dropout to support architecture search with variable operations and edges. SI-VDNAS achieves stable training that would not be affected by the over-selection of skip-connect operation. Experimental results demonstrate that SI-VDNAS finds a convergent architecture with only 2.7 MB parameters within 0.8 GPU-days and can achieve 2.60% top-1 error rate on CIFAR-10. The convergent architecture can obtain a top-1 error rate of 16.20% and 25.6% when transferred to CIFAR-100 and ImageNet (mobile setting).

Federated learning has successfully mitigated the privacy concerns for deep learning over distributed data, but suffers from performance decay led by the non-identical distribution of data. In this paper, we propose a novel framework, namely FedExg, to improve federated learning for non-IID data with model exchange. Two strategies, i.e., FedExg-S and FedExg-A, are developed to realize model exchange in a privacy-preserving fashion. FedExg-S randomly shuffles the models for broadcasting, while FedExg-A improves FedExg-S with partial aggregation to avoid model inversion attack. Theoretical analysis demonstrates that FedExg achieves a reduced upper bound of average regret in comparison to traditional model averaging methods. Experimental results show that FedExg improves image classification tasks on MNIST and CIFAR-10 dataset.

Recently, deep learning based compressive sensing (CS) methods show superior reconstruction performance. However, these approaches are restricted for compressive video sampling in practical design of imaging devices. In this paper, we propose a novel deep network-based CS framework for efficient optical implementation. The proposed framework, dubbed DSCR-Net, consists of a diffractive sensing network employing light diffraction for efficient sampling and a complex-valued neural network for reconstruction, respectively. The diffractive sensing network can achieve real-time sampling at the speed of light. Furthermore, complex-valued neural network is developed to facilitate the reconstruction quality from the complex-valued measurements by jointly considering their real and imaginary parts. Extensive experiments demonstrate that DSCR-Net outperforms the state-of-the-art CS methods in the low sampling-rate region with a potential of immediate implementation of imaging device.

The non-Euclidean geometry characteristic poses a challenge to the saliency prediction for 360-degree images. Since spherical data cannot be projected onto a single plane without distortion, existing saliency prediction methods based on traditional CNNs are inefficient. In this paper, we propose a saliency prediction framework for 360-degree images based on graph convolutional networks (SalGCN), which directly applies to the spherical graph signals. Specifically, we adopt the Geodesic ICOsahedral Pixelation (GICOPix) to construct a spherical graph signal from a spherical image in equirectangular projection (ERP) format. We then propose a graph saliency prediction network to directly extract the spherical features and generate the spherical graph saliency map, where we design an unpooling method suitable for spherical graph signals based on linear interpolation. The network training process is realized by modeling the node regression problem of the input and output spherical graph signals, where we further design a Kullback-Leibler (KL) divergence loss with sparse consistency to make the sparseness of the saliency map closer to the ground truth. Eventually, to obtain the ERP format saliency map for evaluation, we further propose a spherical crown-based (SCB) interpolation method to convert the output spherical graph saliency map into a saliency map in ERP format. Experiments show that our SalGCN can achieve comparable or even better saliency prediction performance both subjectively and objectively, with a much lower computation complexity.

End-to-end optimized image compression has emerged as a disruptive technique to reduce the spatial redundancies with an improved reconstruction quality. However, existing entropy model for latent representations cannot sufficiently exploit their spatial and channel-wise correlations. In this paper, we propose a novel entropy model based on spatial-channel contexts for end-to-end optimized image compression. The proposed model jointly leverages spatial structural dependencies and channel-wise correlations to improve the probabilistic estimation of latent representations. Instead of complex autoregressive hyperprior network, shallow artificial neural networks (ANNs) incorporating 3-D masks are developed to efficiently realize the entropy model with a guarantee of causality. Experimental results demonstrate that the proposed model achieves competitive rate-distortion performance and reduces model complexity in comparison to recent end-to-end optimized methods for image compression.

Conventional scalable compressive video sampling cannot sufficiently exploit the structured sparsity within video sequences. This paper proposes a novel quality scalable structured compressive video sampling (SS-CVS) framework with hierarchical subspace learning to support video transmission over heterogeneous network conditions. The proposed framework incorporates the union of data-driven subspaces (UoDS) model to introduce structured sparsity into sensing matrix for progressive acquisition of measurements. Hierarchical subspace learning is developed to generate the subspaces with their bases based on adaptive subspace clustering in a progressive fashion. Consequently, two hierarchical structures are derived to enable progressive mapping of subspaces and bases on the structured sensing matrices to support quality scalability. To guarantee the convergence of hierarchical subspace learning, the training set is updated with the adaptive group set from the reconstructed reference frames. It is demonstrated that SS-CVS can guarantee stable recovery for each quality layer under the constraints of block restricted isometric property (RIP). Furthermore, the proposed SS-CVS model is generalized to tensor subspaces for scalable compressive sampling of high-dimensional signals. Experimental results demonstrate that the proposed algorithm can achieve quality scalabilities with an improved reconstructed performance in comparison to the state-of-the-art scalable compressive video sampling methods.

Existing light field depth map estimation approaches only utilize partial angular views in occlusion areas and local spatial dependencies in the optimization. This paper proposes a novel two-stage light field depth estimation method via graph spectral analysis to exploit the complete correlations and dependencies within angular patches and spatial images. The initial depth map estimation leverages the undirected graph to jointly consider occluded and unoccluded views within each angular patch. The estimated depth minimizes the structural incoherence of its corresponding angular patch with the focused one by evaluating the highest graph frequency component. Subsequently, depth map refinement optimizes the initial depth map with the color consistency and smoothness formulated by weighted adjacency matrix. The structural constraints are efficiently employed using low-pass graph filtering with Chebyshev polynomial approximation. Experimental results demonstrate that the proposed method improves the depth map estimation, especially in the edge regions.

Network quantization offers an effective solution to deep neural network compression for practical usage. Existing network quantization methods cannot theoretically guarantee the convergence. This paper proposes a novel iterative framework for network quantization with arbitrary bit-widths. We present two Lipschitz constraint based quantization strategies, namely width-level network quantization (WLQ) and multi-level network quantization (MLQ), for high-bit and extremely low-bit (ternary) quantization, respectively. In WLQ, Lipschitz based partition is developed to divide parameters in each layer into two groups: one for quantization and the other for re-training to eliminate the quantization loss. WLQ is further extended to MLQ by introducing layer partition to suppress the quantization loss for extremely low bit-widths. The Lipschitz based partition is proven to guarantee the convergence of the quantized networks. Moreover, the proposed framework is complementary to network compression methods such as activation quantization, pruning and efficient network architectures. The proposed framework is evaluated over extensive state-of-the-art deep neural networks, i.e., AlexNet, VGG-16, GoogleNet and ResNet18. Experimental results show that the proposed framework improves the performance of tasks like classification, object detection and semantic segmentation.

Filter bank based convolutional networks (FBCNs) enable efficient separable multiscale and multidirectional decomposition with a convolutional cascade of 1-D radial and directional filter banks. In this paper, we propose a two-stage subband coding framework for FBCN analysis coefficients using a SPIHT-like algorithm and subsequent primitive-based adaptive arithmetic coding (AAC). The SPIHT-like algorithm extends spatial orientation tree to exploit inter-subband dependency between subbands of different scales and directions. Mutual information is estimated for information-theoretical measurement to formulate such dependencies. Various primitives are designed adaptively encode the generated bitstream by fitting its varying lists and passes. Neural networks are leveraged to improve probability estimation for AAC, where nonlinear prediction is made based on contexts regarding scales, directions, locations and significance of analysis coefficients. Experimental results show that the proposed framework improves the lossy coding performance for FBCN analysis coefficients in comparison to the state-of-the-arts subband coding schemes SPIHT.

The nonlinear recovery is not promising in accuracy and speed, which limits the practical usage of compressed sensing (CS). This paper proposes a deep learning-based CS framework which leverages a deep convolutional auto-encoder for image sensing and recovery. The utilized auto-encoder architecture consists of three components: the fully convolutional network acts as an adaptive measurement matrix generator in the encoder; while in the decoder, the deconvolution network and refined reconstruction network are learned for intermediate and final recovery, respectively. Different from most previous work focusing on the block-wise manner to reduce implementation cost but result in blocky artifacts, our adaptive measurement matrix is applicable to any size of scene image and the decoder network reconstructs the whole image efficiently without any blocky artifacts. Moreover, dense connectivity is leveraged to combine multi-level features and alleviate the vanishing-gradient problem in the refined reconstruction network which boosts the performance on image recovery. Compared to the state-of-the-art methods, our algorithm improves more than 0.8 dB in average PSNR.

Demanding for efficient compression and storage of DNA sequences has been rising with the rapid growth of DNA sequencing technologies. Existing reference-based algorithms map all patterns to regions found in the reference sequence, which lead to redundancy of incomplete similarity. This paper proposes an efficient reference-based method for DNA sequence compression that integrates FM-index and complementary context models to improve compression performance. The proposed method introduces FM-index to represent the full-text matching for exact repeats between the target and reference sequences. For unmatched symbols, complementary context models are leveraged to make weighted estimation conditioned on variable-order contexts. Reversed reference index is used to guarantee the longest match of variable-length substrings. Experimental results show that the proposed method can achieve a 213-fold compression ratio when tested on the first Korean personal genome sequence data set.

Tensor-based compressive sensing (CS) can preserve the intrinsic multidimensional structures with a reduced computational complexity. However, its recovery performance is degraded by simple sparsity. This paper proposes an efficient recovery algorithm GT-ADMM to solve a nonconvex optimization problem for tensor-based CS with group sparsity. Group-sparse representations of tensors is derived based on the Kronecker product of adaptively trained basis. The proposed algorithm incorporates alternating direction method of multipliers (ADMM) to obtain desirable recovery quality with improved efficiency. It requires few degrees of freedom to achieve tensor recovery with a guarantee of minimized approximation error. Experimental results demonstrate that the proposed algorithm outperforms state-of-the-art tensor-based methods in compressive video sampling.

This paper proposes an extended conjugate polar Fourier transform (ECPFT), to design iterated radial filter bank (RFB) and directional filter bank (DFB) convenient for accurate multiscale and multidirectional decomposition in discretization over a convolution network. With conjugated symmetric form, ECPFT would convert complex directional wavelets in original spatial domain to real ones in the inverse Fourier domain of ECPFT. Furthermore, it can contribute to changing the design of nonseparable RFB and DFB to decomposition in scales and directions with 1-D filter bank in each dimension separably in the ECPFT domain. The generated properties of conjugate symmetry and periodicity of 2π in both angular and radial dimensions, also guarantee convolution and sampling operations in both dimensions within the inverse Fourier domain of ECPFT. The fast algorithm for discretization is developed to reduce complexity.

In this paper, we propose a multiscale dictionary learning framework for hierarchical sparse representation of natural images. The proposed framework leverages an adaptive quadtree decomposition to represent structured sparsity in different scales. In dictionary learning, a tree-structured regularized optimization is formulated to distinguish and represent high-frequency details based on varying local statistics and group low-frequency components for local smoothness and structural consistency. In comparison to traditional proximal gradient method, block-coordinate descent is adopted to improve the efficiency of dictionary learning with a guarantee of recovery performance. The proposed framework enables hierarchical sparse representation by naturally organizing the trained dictionary atoms in a prespecified arborescent structure with descending scales from root to leaves. Consequently, the approximation of high-frequency details can be improved with progressive refinement from coarser to finer scales. Employed into image denoising, the proposed framework is demonstrated to be competitive with the state-of-the-art methods in terms of objective and visual restoration quality.

Dictionary learning has emerged as a promising alternative to the conventional hybrid coding framework. However, the rigid structure of sequential training and prediction degrades its performance in scalable video coding. This paper proposes a progressive dictionary learning framework with hierarchical predictive structure for scalable video coding, especially in low bitrate region. For pyramidal layers, sparse representation based on spatio-temporal dictionary is adopted to improve the coding efficiency of enhancement layers with a guarantee of reconstruction performance. The overcomplete dictionary is trained to adaptively capture local structures along motion trajectories as well as exploit the correlations between the neighboring layers of resolutions. Furthermore, progressive dictionary learning is developed to enable the scalability in temporal domain and restrict the error propagation in a closed-loop predictor. Under the hierarchical predictive structure, online learning is leveraged to guarantee the training and prediction performance with an improved convergence rate. To accommodate with the state-of-the-art scalable extension of H.264/AVC and latest High Efficiency Video Coding (HEVC), standardized codec cores are utilized to encode the base and enhancement layers. Experimental results show that the proposed method outperforms the latest scalable extension of HEVC and HEVC simulcast over extensive test sequences with various resolutions.

Conventional Compressive Sensing (CS) obscures the intrinsic structures of multidimensional signals with the vectorized representation. Although tensor-based CS methods can preserve the intrinsic multidimensional structures with reduced computational complexity, their sampling efficiency and recovery performance are degraded with the assumption of standard/simple sparsity. This paper proposes a general and adaptive model that incorporates structured sparsity into tensor representation to fit the varying nonstationary statistics of multidimensional signals. To guarantee the block sparsity, subspace clustering is adopted to adaptively generate the union of tensor subspaces with its basis of each tensor subspace learned for optimized representation. For sampled tensors, the stable recovery algorithm is developed to achieve desirable recovery performance using fewer degrees of freedom. Moreover, the proposed model inherits the merit from tensor-based CS to alleviate the computational and storage burden in sampling and recovery. Experimental results demonstrate that the proposed model can achieve better recovery performance in video sampling in comparison to the state-of-the-art tensor-based method.

MOTIVATION: Genome-wide association studies (GWAS) have been widely used in discovering the association between genotypes and phenotypes. Human genome data contain valuable but highly sensitive information. Unprotected disclosure of such information might put individual's privacy at risk. It is important to protect human genome data. Exact logistic regression is a bias-reduction method based on a penalized likelihood to discover rare variants that are associated with disease susceptibility. We propose the HEALER framework to facilitate secure rare variants analysis with a small sample size. RESULTS: We target at the algorithm design aiming at reducing the computational and storage costs to learn a homomorphic exact logistic regression model (i.e. evaluate P-values of coefficients), where the circuit depth is proportional to the logarithmic scale of data size. We evaluate the algorithm performance using rare Kawasaki Disease datasets. AVAILABILITY AND IMPLEMENTATION: Download HEALER at http://research.ucsd-dbmi.org/HEALER/ CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Classical dictionary learning methods for video coding suffer from high computational complexity and interfered coding efficiency by disregarding its underlying distribution. This paper proposes a spatio-temporal online dictionary learning (STOL) algorithm to speed up the convergence rate of dictionary learning with a guarantee of approximation error. The proposed algorithm incorporates stochastic gradient descents to form a dictionary of pairs of 3D low-frequency and high-frequency spatio-temporal volumes. In each iteration of the learning process, it randomly selects one sample volume and updates the atoms of dictionary by minimizing the expected cost, rather than optimizes empirical cost over the complete training data, such as batch learning methods, e.g., K-SVD. Since the selected volumes are supposed to be independent identically distributed samples from the underlying distribution, decomposition coefficients attained from the trained dictionary are desirable for sparse representation. Theoretically, it is proved that the proposed STOL could achieve better approximation for sparse representation than K-SVD and maintain both structured sparsity and hierarchical sparsity. It is shown to outperform batch gradient descent methods (K-SVD) in the sense of convergence speed and computational complexity, and its upper bound for prediction error is asymptotically equal to the training error. With lower computational complexity, extensive experiments validate that the STOL-based coding scheme achieves performance improvements than H.264/AVC or High Efficiency Video Coding as well as existing super-resolution-based methods in rate-distortion performance and visual quality.

Summary form only given. This paper proposes a bi-directional context modeling (BCM) technique for reference-free genome sequence compression, which constructs its contexts by combining arbitrary predicted symbols in two directions corresponding to approximate repeats and non-repeat regions. Thus, BCM can sequentially predict DNA sequences with weighted conditional probabilities that simultaneously exploit the correlations among matched approximate repeats and fit the variable-order statistics in non-repeat regions. Moreover, BCM eliminates the overhead of pointer information for specifying approximate repeats, as it is synchronized in both encoder and decoder. In theory, we show that upper bounds of excess model redundancy led by BCM vanish with the growth of sequence size. Experimental results show that BCM outperforms the state-of-the-art reference-free compressors like FCM and CTW+LZ.

Existing sparse representation with subspace learning is hampered by the intersection of subspaces of bases. With structured sparsity to enable the prior knowledge of signal statistics, this paper proposes a novel compressive video sampling by subspace learning to minimize the intersection of subspaces. As the measurement, the block coherence is optimized with the regularized learning to generate a class of independent bases associated with the subspaces. Thus, the proposed framework can make a compact block sparse representation based on the derived basis in an efficient and adaptive manner. The block-based recovery of video sequences is demonstrated to be stable under the constraint of block restricted isometric property (RIP). Experimental results show that the proposed method outperforms existing compressive video sampling schemes.

This paper proposes an adaptive dictionary learning approach based on sub modular optimization. A candidate atom set is constructed based on multiple bases from the combination of analytic and trained dictionaries. With the low-frequency components by the analytic DCT atoms, high-resolution dictionaries can be inferred through online learning to make efficient approximation with rapid convergence. It is formulated as a combinatorial optimization for approximate sub modularity, which is suitable for sparse representation based on dictionaries with arbitrary structures. In single-image super-resolution, the proposed scheme has been demonstrated to improve the reconstruction performance in comparison with double sparsity dictionary in terms of both objective and subjective restoration quality.

To enable learning-based video coding for transmission over heterogenous networks, this paper proposes a scalable video coding framework by progressive dictionary learning. With the hierarchical B-picture prediction structure, the inter-predicted frames would be reconstructed in terms of the spatio-temporal dictionary in a successive sense. Within the progressive dictionary learning, the training set is enriched with the samples from the reconstructed frames in the coarse layer. Through minimizing the expected cost, the stochastic gradient descent is leveraged to update the dictionary for practical coding. It is demonstrated that the learning-based scalable framework can effectively guarantee the consistency of motion trajectory with the well-designed spatio-temporal dictionary.

LS-based adaptation cannot fully exploit high-dimensional correlations in image signals, as linear prediction model in the input space of supports is undesirable to capture higher order statistics. This paper proposes Gaussian process regression for prediction in lossless image coding. Incorporating kernel functions, the prediction support is projected into a high-dimensional feature space to fit the anisotropic and nonlinear image statistics. Instead of directly conditioned on the support, Gaussian process regression is leveraged to make prediction in the feature space. The model parameters are optimized by measuring the similarities based on the training set, which is evaluated by combined kernel function in the sense of translation and rotation invariance among supports mapped in the feature space. Experimental results show that the proposed predictor outperforms most benchmark predictors reported.

Inherent statistical correlation for context-based prediction and structural interdependencies for local coherence is not fully exploited in existing lossless image coding schemes. This paper proposes a novel prediction model where the optimal correlated prediction for a set of pixels is obtained in the sense of the least code length. It not only exploits the spatial statistical correlations for the optimal prediction directly based on 2D contexts, but also formulates the data-driven structural interdependencies to make the prediction error coherent with the underlying probability distribution for coding. Under the joint constraints for local coherence, max-margin Markov networks are incorporated to combine support vector machines structurally to make max-margin estimation for a correlated region. Specifically, it aims to produce multiple predictions in the blocks with the model parameters learned in such a way that the distinction between the actual pixel and all possible estimations is maximized. It is proved that, with the growth of sample size, the prediction error is asymptotically upper bounded by the training error under the decomposable loss function. Incorporated into the lossless image coding framework, the proposed model outperforms most prediction schemes reported.

Previous reference-based compression on DNA sequences do not fully exploit the intrinsic statistics by merely concerning the approximate matches. In this paper, an adaptive difference distribution-based coding framework is proposed by the fragments of nucleotides with a hierarchical tree structure. To keep the distribution of difference sequence from the reference and target sequences concentrated, the sub-fragment size and matching offset for predicting are flexible to the stepped size structure. The matching with approximate repeats in reference will be imposed with the Hamming-like weighted distance measure function in a local region closed to the current fragment, such that the accuracy of matching and the overhead of describing matching offset can be balanced. A well-designed coding scheme will make compact both the difference sequence and the additional parameters, e.g. sub-fragment size and matching offset. Experimental results show that the proposed scheme achieves 150% compression improvement in comparison with the best reference-based compressor GReEn.

In this paper, we develop a novel genome compression framework based on distributed source coding (DSC)[3], which is specially tailored to the need of miniaturized devices. At the encoder side, subsequences with adaptive code length can be compressed flexibly through either low complexity DSC based syndrome coding or hash coding with the decision determined by the existence of variations between source and reference known from the decoder feedback. Moreover, to tackle the variations between source and reference at the decoder, we carefully designed a factor graph based low-density parity-check (LDPC) decoder, which automatically detects insertion, deletion and substitution.

This paper proposes a novel model on intra coding for High Efficiency Video Coding (HEVC), which simultaneously predicts blocks of pixels with optimal rate distortion. It utilizes the spatial statistical correlation for the optimal prediction based on 2-D contexts, in addition to formulating the data-driven structural interdependences to make the prediction error coherent with the probability distribution, which is desirable for successful transform and coding. The structured set prediction model incorporates a max-margin Markov network (M3N) to regulate and optimize multiple block predictions. The model parameters are learned by discriminating the actual pixel value from other possible estimates to maximize the margin (i.e., decision boundary bandwidth). Compared to existing methods that focus on minimizing prediction error, the M3N-based model adaptively maintains the coherence for a set of predictions. Specifically, the proposed model concurrently optimizes a set of predictions by associating the loss for individual blocks to the joint distribution of succeeding discrete cosine transform coefficients. When the sample size grows, the prediction error is asymptotically upper bounded by the training error under the decomposable loss function. As an internal step, we optimize the underlying Markov network structure to find states that achieve the maximal energy using expectation propagation. For validation, we integrate the proposed model into HEVC for optimal mode selection on rate-distortion optimization. The proposed prediction model obtains up to 2.85% bit rate reduction and achieves better visual quality in comparison to the HEVC intra coding.

In this paper, we investigate and propose a novel prediction model for lossless image coding in which the optimal correlated prediction for block of pixels are simultaneously obtained in the sense of the least code length. It not only utilizes the spatial statistical correlation for the optimal prediction directly based on 2-D contexts, but also formulates the data-driven structural interdependencies to make the prediction error coherent with the underlying probability distribution for coding. Besides the discriminative adaptive pixel-wise prediction, the Markov network is adaptively derived to maintain the coherence of prediction in the blocks and seek the concurrent optimization of set of prediction by relating the loss function to actual code length. The prediction error is shown to be asymptotically upper bounded by the training error under the decomposable loss function. For validation, we apply the proposed model into lossless image coding and experimental results show that the proposed scheme outperforms the best prediction scheme reported.

This paper proposes a novel model on intra-coding for high efficiency video coding (HEVC), which can simultaneously make the set of prediction for block of pixels in an optimal rate-distortion sense. It not only utilizes the spatial statistical correlation for the optimal prediction based on 2-D contexts, but also formulates the data-driven structural interdependencies to make the prediction error coherent with the probability distribution which is favorable for subsequent transform and coding. The so-called structured set prediction model incorporates max-margin Markov network to regulate and reason the multiple prediction in the blocks. The model parameters are learned by discriminating the actual pixel value from the other possible estimates to the maximal margin. Distinguished from the existing methods concerning the minimal prediction error, the Markov network is adaptively derived to maintain the coherence of set of prediction. To be concrete, the proposed model seeks the concurrent optimization of the set of prediction by relating the loss function to the probability distribution of subsequent DCT coefficients. The prediction error is demonstrated to be asymptotically upper bounded by the training error under the decomposable loss function. For validation, we integrate the proposed model into HEVC intra coding and experimental results show obvious improvement of coding performance in terms of BD-rate.

In this paper, we propose the bitwise structured prediction model for lossless image coding, especially for the oscillatory regions. The learning-based model utilizes the regular features obtained from the predicted local data. At first, the pixel-wise prediction is decomposed into the bitwise ones. In each bit plane, the prediction of the current bit is simplified to the max margin estimation for the 0/1 prediction problem and obtained directly conditioned on the neighboring predicted bits. Furthermore, since the decreasing dependencies of neighboring bits in lower bit plane lead to the turbulence of predictive results, the structured prediction is proposed to establish the Markov network to constrain the outputs of the bit planes, and suppress the prediction errors with a well-defined loss function. Consequently, the min-max formulation is proposed for the concurrent optimization for maximizing the 0/1 margin of all the bit planes.

The sequential context modeling framework is generalized to a non-sequential one by context relaxation from consecutive suffix of the subsequences of symbols to the permutation of the preceding symbols as result of considering complex context structures in such sources as video and program binaries. Context weighting tree is also extended to a series of context trees which are built according to the "model tree", in which the descendent relationship in the formation of non-sequential context sets is described. Model redundancy and maximum a posteriori model in the framework are discussed and compared. A decision method based on the greedy algorithm is proposed to customize sets of models fitting the concrete sources. Brief description of application to executable data files incorporating with the semantics and syntax constraints are given and experiment are made accordingly as a validation.

The use of sensors in industrial welding productions shall be undoubtedly advantageous to enhance the quality, flexibility and efficiency of robots. However sensors are only integrated in a small number of applications. An essential reason is the difficulty to generate robot programs for sensor operations. In this paper a new method for off-line programming of sensor operations is described. The developed module has been integrated in a PC-based off-line programming system (called PIN). Its user-friendly design enables fast and easy programming of sensor operations assisted by icon-oriented user interface and task-sorted macro library. In spite of running on a PC (low-cost system), a high performance is offered by the presented system which was developed for the use of nonprogramming specialists.