Novel object synthesis by integrating distinct textual concepts from diverse categories remains a significant challenge in text-to-image generation. Existing methods often suffer from insufficient concept mixing, lack of rigorous evaluation, and suboptimal outputs, resulting in conceptual imbalance, superficial combinations, or mere juxtapositions. To address these limitations, we propose Reinforcement Mixing Learning (RMLer), a framework that formulates cross-category concept fusion as a reinforcement learning problem: mixed features serve as states, mixing strategies as actions, and visual outcomes as rewards. Specifically, we design an MLP policy network to predict dynamic coefficients for blending cross-category text embeddings. We further introduce visual rewards based on (1) semantic similarity and (2) compositional balance between the fused object and its constituent concepts, and optimize the policy via proximal policy optimization. At inference time, a selection strategy leverages these rewards to curate the highest-quality fused objects. Extensive experiments demonstrate that RMLer synthesizes coherent, high-fidelity objects from diverse categories and consistently outperforms existing methods. Our work provides a robust framework for generating novel visual concepts, with promising applications in film, gaming, and design.

Intelligent systems typically need to continually learn from streaming data subject to distribution shift, where a key requirement is that they cannot catastrophically forget the historical knowledge learned from previous data. More seriously, streaming data often contain substantial label noise, which can exacerbate catastrophic forgetting and lead to performance degradation on forthcoming data. To address these problems, Continual Noisy Label Learning (CNLL) has been proposed. However, existing CNLL methods still fall short of the ability in addressing catastrophic forgetting because they adopted heuristic strategies in handling label noise and did not explicitly characterize the distributional shift across time, which hinders effective knowledge transfer from historical data to new data. To tackle these challenges, we theoretically analyze the problem of learning from streaming noisy data with distribution shift and propose a unified framework called Continual Noisy Label Learning on Drifting Data Streams (CNLDD). Specifically, we theoretically explore, for the first time, the upper bound of cumulative generalization error for CNLL problem, which reveals three factors leading to forgetting, namely selection bias of buffered data, distribution shift, and label noise. To alleviate the selection bias of buffered data, we design a two-step buffer update strategy to narrow the distribution gap between the original historical data and the selected representative data in buffer. To address distribution shift, our CNLDD explicitly characterizes the distribution discrepancies between buffered data and incoming data, prioritizing historical data with minimal discrepancies to enhance knowledge transfer. To tackle noisy labels, CNLDD estimates the importance weight of each example with the instance-dependent noise transition matrix, thereby avoiding the data bias and knowledge forgetting arising from noisy labels. Empirically, due to the unified modeling of the aforementioned issues, our CNLDD achieves superior classification performance when compared with state-of-the-art CNLL methods on both synthetic and real-world datasets.

Diffusion models (DMs) have been successfully applied to real image editing. These models typically invert images into latent noise vectors during the inversion process, and then edit them during the inference process. However, DMs often rely on the local linearization assumption, which assumes that the noise injected during the inversion process approximates the noise removed during the inference process. While DMs efficiently generate images under this assumption, it also accumulates errors during the diffusion process due to the assumption, ultimately negatively impacting the quality of real image reconstruction and editing. To address this issue, we propose a novel ERDDCI (Exact Reversible Diffusion via Dual-Chain Inversion). ERDDCI uses the new Dual-Chain Inversion (DCI) for joint inference to derive an exact reversible diffusion process. Using DCI, our method avoids the cumbersome optimization process in existing inversion approaches and achieves high-quality image editing. Additionally, to accommodate image operations under high guidance scales, we introduce a dynamic control strategy that enables more refined image reconstruction and editing. Our experiments demonstrate that ERDDCI significantly outperforms state-of-the-art methods in a 50-step diffusion process. It achieves rapid and precise image reconstruction with SSIM of 0.999 and LPIPS of 0.001, and delivers competitive results in image editing. The source code is available at: https://github.com/daii-y/ERDDCI.

Recently, Graph Transformer (GT) models have been widely used in the task of Molecular Property Prediction (MPP) due to their high reliability in characterizing the latent relationship among graph nodes (i.e., the atoms in a molecule). However, most existing GT-based methods usually explore the basic interactions between pairwise atoms, and thus they fail to consider the important interactions among critical motifs (e.g., functional groups consisted of several atoms) of molecules. As motifs in a molecule are significant patterns that are of great importance for determining molecular properties (e.g., toxicity and solubility), overlooking motif interactions inevitably hinders the effectiveness of MPP. To address this issue, we propose a novel Atom-Motif Contrastive Transformer (AMCT), which not only explores the atom-level interactions but also considers the motif-level interactions. Since the representations of atoms and motifs for a given molecule are actually two different views of the same instance, they are naturally aligned to generate the self-supervisory signals for model training. Meanwhile, the same motif can exist in different molecules, and hence we also employ the contrastive loss to maximize the representation agreement of identical motifs across different molecules. Finally, in order to clearly identify the motifs that are critical in deciding the properties of each molecule, we further construct a property-aware attention mechanism into our learning framework. Our proposed AMCT is extensively evaluated on 10 popular benchmark datasets, and both quantitative and qualitative results firmly demonstrate its effectiveness when compared with the state-of-the-art methods.

Data-free knowledge distillation aims to learn a compact student network from a pre-trained large teacher network without using the original training data of the teacher network. Existing collection-based and generation-based methods train student networks by collecting massive real examples and generating synthetic examples, respectively. However, they inevitably become weak in practical scenarios due to the difficulties in gathering or emulating sufficient real-world data. To solve this problem, we propose a novel method called Hybrid Data-Free Distillation (HiDFD), which leverages only a small amount of collected data as well as generates sufficient examples for training student networks. Our HiDFD comprises two primary modules, i.e., the teacher-guided generation and student distillation. The teacher-guided generation module guides a Generative Adversarial Network (GAN) by the teacher network to produce high-quality synthetic examples from very few real-world collected examples. Specifically, we design a feature integration mechanism to prevent the GAN from overfitting and facilitate the reliable representation learning from the teacher network. Meanwhile, we drive a category frequency smoothing technique via the teacher network to balance the generative training of each category. In the student distillation module, we explore a data inflation strategy to properly utilize a blend of real and synthetic data to train the student network via a classifier-sharing-based feature alignment technique. Intensive experiments across multiple benchmarks demonstrate that our HiDFD can achieve state-of-the-art performance using 120 times less collected data than existing methods.

Heterogeneity is a fundamental and challenging issue in federated learning, especially for the graph data due to the complex relationships among the graph nodes. To deal with the heterogeneity, lots of existing methods perform the weighted federation based on their calculated similarities between pairwise clients (i.e., subgraphs). However, their inter-subgraph similarities estimated with the outputs of local models are less reliable, because the final outputs of local models may not comprehensively represent the real distribution of subgraph data. In addition, they ignore the critical intra-heterogeneity which usually exists within each subgraph itself. To address these issues, we propose a novel Federated learning method by integrally modeling the Inter-Intra Heterogeneity (FedIIH). For the inter-subgraph relationship, we propose a novel hierarchical variational model to infer the whole distribution of subgraph data in a multi-level form, so that we can accurately characterize the inter-subgraph similarities with the global perspective. For the intra-heterogeneity, we disentangle the subgraph into multiple latent factors and partition the model parameters into multiple parts, where each part corresponds to a single latent factor. Our FedIIH not only properly computes the distribution similarities between subgraphs, but also learns disentangled representations that are robust to irrelevant factors within subgraphs, so that it successfully considers the inter- and intra- heterogeneity simultaneously. Extensive experiments on six homophilic and five heterophilic graph datasets in both non-overlapping and overlapping settings demonstrate the effectiveness of our method when compared with eight state-of-the-art methods. Specifically, FedIIH averagely outperforms the second-best method by a large margin of 5.79% on all heterophilic datasets.

Spherical Sliced-Wasserstein (SSW) has recently been proposed to measure the discrepancy between spherical data distributions in various fields, such as geology, medical domains, computer vision, and deep representation learning. However, in the original SSW, all projection directions are treated equally, which is too idealistic and cannot accurately reflect the importance of different projection directions for various data distributions. To address this issue, we propose a novel data-adaptive Discriminative Spherical Sliced-Wasserstein (DSSW) distance, which utilizes a projected energy function to determine the discriminative projection direction for SSW. In our new DSSW, we introduce two types of projected energy functions to generate the weights for projection directions with complete theoretical guarantees. The first type employs a non-parametric deterministic function that transforms the projected Wasserstein distance into its corresponding weight in each projection direction. This improves the performance of the original SSW distance with negligible additional computational overhead. The second type utilizes a neural network-induced function that learns the projection direction weight through a parameterized neural network based on data projections. This further enhances the performance of the original SSW distance with less extra computational overhead. Finally, we evaluate the performance of our proposed DSSW by comparing it with several state-of-the-art methods across a variety of machine learning tasks, including gradient flows, density estimation on real earth data, and self-supervised learning.

Physiological activities can be manifested by the sensitive changes in facial imaging. While they are barely observable to our eyes, computer vision manners can, and the derived remote photoplethysmography (rPPG) has shown considerable promise. However, existing studies mainly rely on spatial skin recognition and temporal rhythmic interactions, so they focus on identifying explicit features under ideal light conditions, but perform poorly in-the-wild with intricate obstacles and extreme illumination exposure. In this paper, we propose an end-to-end video transformer model for rPPG. It strives to eliminate complex and unknown external time-varying interferences, whether they are sufficient to occupy subtle biosignal amplitudes or exist as periodic perturbations that hinder network training. In the specific implementation, we utilize global interference sharing, subject background reference, and self-supervised disentanglement to eliminate interference, and further guide learning based on spatiotemporal filtering, reconstruction guidance, and frequency domain and biological prior constraints to achieve effective rPPG. To the best of our knowledge, this is the first robust rPPG model for real outdoor scenarios based on natural face videos, and is lightweight to deploy. Extensive experiments show the competitiveness and performance of our model in rPPG prediction across datasets and scenes.

Time Series Forecasting (TSF) is critical in many real-world domains like financial planning and health mon-itoring. Recent studies have revealed that Large Language Models (LLMs), with their powerful in-contextual modeling capabilities, hold significant potential for TSF. However, existing LLM-based methods usually perform suboptimally because they neglect the inherent characteristics of time series data. Unlike the textual data used in LLM pre-training, the time series data is semantically sparse and comprises distinctive tempo-ral patterns. To address this problem, we propose LLM-PS to empower the LLM for TSF by learning the fundamental Patterns and meaningful Semantics from time series data. Our LLM-PS incorporates a new multi-scale convolutional neural network adept at capturing both short-term fluctuations and long-term trends within the time series. Meanwhile, we introduce a time-to-text module for extracting valuable semantics across continuous time intervals rather than isolated time points. By integrating these patterns and semantics, LLM - PS effectively models temporal dependencies, enabling a deep comprehension of time series and delivering accurate forecasts. Intensive exper-imental results demonstrate that LLM-PS achieves state-of-the-art performance in both short- and long-term forecasting tasks, as well as in few- and zero-shot settings. Code is available at https://github.com/tangjialiang97ILLMPS.

Measuring the similarity between data points plays a vital role in lots of popular representation learning tasks such as metric learning and contrastive learning. Most existing approaches utilize point-level distances to learn the point-to-point similarity between pairwise instances. However, since the finite number of training data points cannot fully cover the whole sample space consisting of an infinite number of points, the generalizability of the learned distance is usually limited by the sample size. In this paper, we thus extend the conventional form of data point to the new form of data ball with a predictable volume, so that we can naturally generalize the existing point-level distance to a new volume-aware distance (VAD) which measures the field-to-field geometric similarity. The learned VAD not only takes into account the relationship between observed instances but also uncovers the similarity among those unsampled neighbors surrounding the training data. This practice significantly enriches the coverage of sample space and thus improves the model generalizability. Theoretically, we prove that VAD tightens the error bound of traditional similarity learning and preserves crucial topological properties. Experiments on multi-domain data demonstrate the superiority of VAD over existing approaches in both supervised and unsupervised tasks.

Data augmentation has been widely used in machine learning. Its main goal is to transform and expand the original data using various techniques, creating a more diverse and enriched training dataset. However, due to the disorder and irregularity of point clouds, existing methods struggle to enrich geometric diversity and maintain topological consistency, leading to imprecise point cloud understanding. In this paper, we propose SinPoint, a novel method designed to preserve the topological structure of the original point cloud through a homeomorphism. It utilizes the Sine function to generate smooth displacements. This simulates object deformations, thereby producing a rich diversity of samples. In addition, we propose a Markov chain Augmentation Process to further expand the data distribution by combining different basic transformations through a random process. Our extensive experiments demonstrate that our method consistently outperforms existing Mixup and Deformation methods on various benchmark point cloud datasets, improving performance for shape classification and part segmentation tasks. Specifically, when used with PointNet++ and DGCNN, our method achieves a state-of-the-art accuracy of 90.2 in shape classification with the real-world ScanObjectNN dataset. We release the code at https://github.com/CSBJian/SinPoint.

Multi-label learning (MLL) has gained attention for its ability to represent real-world data. Label Distribution Learning (LDL), an extension of MLL to learning from label distributions, faces challenges in collecting accurate label distributions. To address the issue of biased annotations, based on the low-rank assumption, existing works recover true distributions from biased observations by exploring the label correlations. However, recent evidence shows that the label distribution tends to be full-rank, and naive apply of low-rank approximation on biased observation leads to inaccurate recovery and performance degradation. In this paper, we address the LDL with biased annotations problem from a novel perspective, where we first degenerate the soft label distribution into a hard multi-hot label and then recover the true label information for each instance. This idea stems from an insight that assigning hard multi-hot labels is often easier than assigning a soft label distribution, and it shows stronger immunity to noise disturbances, leading to smaller label bias. Moreover, assuming that the multi-label space for predicting label distributions is low-rank offers a more reasonable approach to capturing label correlations. Theoretical analysis and experiments confirm the effectiveness and robustness of our method on real-world datasets.

Contrastive learning has achieved remarkable success in self-supervised learning by pretraining a generalizable feature representation based on the augmentation invariance. Most existing approaches assume that different augmented views of the same instance (i.e., the *positive pairs*) remain semantically invariant. However, the augmentation results with *varying extent* may introduce semantic discrepancies or even content distortion, and thus the conventional (pseudo) supervision from augmentation invariance may lead to misguided learning objectives. In this paper, we propose a novel method called Contrastive Learning with Variable Similarity (CLVS) to accurately characterize the intrinsic similarity relationships between different augmented views. Our method dynamically adjusts the similarity based on the augmentation extent, and it ensures that strongly augmented views are always assigned lower similarity scores than weakly augmented ones. We provide a theoretical analysis to guarantee the effectiveness of the variable similarity in improving model generalizability. Extensive experiments demonstrate the superiority of our approach, achieving gains of 2.1\% on ImageNet-100 and 1.4\% on ImageNet-1k compared with the state-of-the-art methods.

Pseudo label based semi-supervised learning (SSL) for single-label and multi-label classification tasks has been extensively studied; however, semi-supervised label distribution learning (SSLDL) remains a largely unexplored area. Existing SSL methods fail in SSLDL because the pseudo-labels they generate only ensure overall similarity to the ground truth but do not preserve the ranking relationships between true labels, as they rely solely on KL divergence as the loss function during training. These skewed pseudo-labels lead the model to learn incorrect semantic relationships, resulting in reduced performance accuracy. To address these issues, we propose a novel SSLDL method called \textit{RankMatch}. \textit{RankMatch} fully considers the ranking relationships between different labels during the training phase with labeled data to generate higher-quality pseudo-labels. Furthermore, our key observation is that a flexible utilization of pseudo-labels can enhance SSLDL performance. Specifically, focusing solely on the ranking relationships between labels while disregarding their margins helps prevent model overfitting. Theoretically, we prove that incorporating ranking correlations enhances SSLDL performance and establish generalization error bounds for \textit{RankMatch}. Finally, extensive real-world experiments validate its effectiveness.

Real-world data may contain a considerable amount of noisily labeled examples, which usually mislead the training algorithm and result in degraded classification performance on test data. Therefore, Label Noise Learning (LNL) was proposed, of which one popular research trend focused on estimating the critical statistics (e.g., sample mean and sample covariance), to recover the clean data distribution. However, existing methods may suffer from the unreliable sample selection process or can hardly be applied to multi-class cases. Inspired by the centroid estimation theory, we propose Per-Class Statistic Estimation (PCSE), which establishes the quantitative relationship between the clean (first-order and second-order) statistics and the corresponding noisy statistics for every class. This relationship is further utilized to induce a generative classifier for model inference. Unlike existing methods, our approach does not require sample selection from the instance level. Moreover, our PCSE can serve as a general post-processing strategy applicable to various popular networks pre-trained on the noisy dataset for boosting their classification performance. Theoretically, we prove that the estimated statistics converge to their ground-truth values as the sample size increases, even if the estimated label transition matrix is biased. Empirically, we conducted intensive experiments on various binary and multi-class datasets, and the results demonstrate that PCSE achieves more precise statistic estimation as well as higher classification accuracy when compared with state-of-the-art methods in LNL.

Audio-Visual Segmentation (AVS) aims to segment sound sources from video frames using synchronized audio cues. This task requires not only localizing the sound sources within frames but also accurately delineating their shapes. Existing AVS methods often rely on assumptions of spatial-temporal consistency between audio-visual content and are typically designed for specific learning paradigms. However, this specialization limits their ability to handle multi-granularity supervision signals and and adapt to diverse task requirements. For this purpose, we propose a Cross-modal Spatial-Temporal Alignment (X-STA) network to alleviate spatial-temporal inconsistency and overcome paradigm-specific constraint. Our X-STA introduces three key components: a novel multi-stage Cross-modal Adapter (xAdapter) that transfers knowledge from a pre-trained SAM through multi-grained representation adaptation, an innovative Cross-modal Prompter (xPrompter) that provides geometry-aware constraints for AVS through dynamic prompting strategies, and a Cross-modal Self-supervised (xSelf) mechanism that refines temporal alignment and enables self-supervised AVS. These components collectively facilitate explicit reasoning about location and geometric shape of the sound source by refining the alignment of cross-modal spatial-temporal cues. Our method achieves competitive performance across several baselines on widely-used AVS datasets, demonstrating its effectiveness in addressing the complexities of AVS.

Feature noise and label noise are ubiquitous in practical scenarios, which pose great challenges for training a robust machine learning model. Most previous approaches usually deal with only a single problem of either feature noise or label noise. However, in real-world applications, hybrid noise, which contains both feature noise and label noise, is very common due to the unreliable data collection and annotation processes. Although some results have been achieved by a few representation learning based attempts, this issue is still far from being addressed with promising performance and guaranteed theoretical analyses. To address the challenge, we propose a novel unified learning framework called Feature and Label Recovery (FLR) to combat the hybrid noise from the perspective of data recovery, where we concurrently reconstruct both the feature matrix and the label matrix of input data. Specifically, the clean feature matrix is discovered by the low-rank approximation, and the ground-truth label matrix is embedded based on the recovered features with a nuclear norm regularization. Meanwhile, the feature noise and label noise are characterized by their respective adaptive matrix norms to satisfy the corresponding maximum likelihood. As this framework leads to a non-convex optimization problem, we develop the non-convex Alternating Direction Method of Multipliers (ADMM) with the convergence guarantee to solve our learning objective. We also provide the theoretical analysis to show that the generalization error of FLR can be upper-bounded in the presence of hybrid noise. Experimental results on several typical benchmark datasets clearly demonstrate the superiority of our proposed method over the state-of-the-art robust learning approaches for various noises.

Recently, multi-view subspace clustering has attracted extensive attention due to the rapid increase of multi-view data in many real-world applications. The main goal of this task is to learn a common representation of multiple subspaces from the given multi-view data, and most existing methods usually directly merge multiple groups of features by the single-step integration. However, there may exist large disparities among different views of the data, and thus the conventional single-step practice can hardly obtain a generally consistent feature representation for the multi-view data. To overcome this challenge, we present a novel approach dubbed “Asymptotics-Aware Multi-view Subspace Clustering (A$^{2}$MSC)” to pursue a consistent feature representation in a multi-step way, which iteratively conducts the data recovery to gradually reduce the differences between pairwise views. Specifically, we construct an asymptotic learning rule to update the feature representation, and the iteration result converges to a consistent feature vector for characterizing each instance of the original multi-view data. After that, we utilize such a new feature representation to learn a clustering-oriented similarity matrix via minimizing a self-expressive objective, and we also design the corresponding optimization algorithm to solve it with convergence guarantees. Theoretically, we prove that the learned asymptotic representation effectively integrates multiple views, thereby ensuring the effective handling of multi-view data. Empirically, extensive experimental results demonstrate the superiority of our proposed A$^{2}$MSC over the state-of-the-art multi-view subspace clustering approaches.

Self-supervised contrastive learning (CL) seeks to learn generalizable feature representations via the self-supervision of pairwise similarities, where existing CL approaches usually build definite similarity labels (e.g., positive or negative) for model training. Yet in practice, the same pair of instances may have opposite similarity labels in different scenarios, e.g., two interclass images from CIFAR-100 can be a similar pair in CIFAR-20. Learning with definite similarities can hardly obtain an ideal representation that simultaneously characterizes the similar and dissimilar patterns (e.g., the contexts and details) between each two instances. Therefore, pairwise similarities used for CL should be agnostic, and we argue that simultaneously considering both the similarity and dissimilarity for each data pair could learn more generalizable representations. To this end, we propose similarity-agnostic CL (SACL), which generalizes the instance discrimination strategy of conventional CL to a new multiobjective programming (MOP) form. In SACL, we build multiple projection layers with corresponding regularizers to constrain the distance matrix to have different sparsity in different objectives so that we can obtain alterable pairwise distances to capture both the similarity and dissimilarity between each pair of instances. We show that SACL can be equivalently converted to a single learning objective, easily solved by stochastic optimization with convergence guarantees. Theoretically, we prove a tighter error bound than conventional CL approaches; empirically, our method improves the downstream task performance for image, text, and graph data.

In zero-shot learning (ZSL), attribute knowledge plays a vital role in transferring knowledge from seen classes to unseen classes. However, most existing ZSL methods learn biased attribute knowledge, which usually results in biased attribute prediction and a decline in zero-shot recognition performance. To solve this problem and learn unbiased attribute knowledge, we propose a visual attribute Transformer for zero-shot recognition (ZS-VAT), which is an effective and interpretable Transformer designed specifically for ZSL. In ZS-VAT, we design an attribute-head self-attention (AHSA) that is capable of learning unbiased attribute knowledge. Specifically, each attribute head in AHSA first transforms the local features into attribute-reinforced features and then accumulates the attribute knowledge from all corresponding reinforced features, reducing the mutual influence between attributes and avoiding information loss. AHSA finally preserves unbiased attribute knowledge through attribute embeddings. We also propose an attribute fusion model (AFM) that learns to recover the correct category knowledge from the attribute knowledge. In particular, AFM takes all features from AHSA as input and generates global embeddings. We carried out experiments to demonstrate that the attribute knowledge from AHSA and the category knowledge from AFM are able to assist each other. During the final semantic prediction, we combine the attribute embedding prediction (AEP) and global embedding prediction (GEP). We evaluated the proposed scheme on three benchmark datasets. ZS-VAT outperformed the state-of-the-art generalized ZSL (GZSL) methods on two datasets and achieved competitive results on the other dataset.

The self-expressive strategy has shown excellent capabilities in realizing low-dimensional representations of high-dimensional data for subspace clustering algorithms. The existing designs, however, are formulated on the linearization assumptions of the data, neglecting the precise characterization of linear relationships within samples. Considering that real-world data adheres to diverse distribution forms, it becomes impractical to first treat the samples as existing in a uniform linear space before finding an appropriate manifold space. To handle this challenge, we propose a novel self-expressive-based learning framework termed metric learning-based subspace clustering (MLSC). Particularly, we smoothly incorporate metric learning into the subspace clustering framework by introducing adaptive neighbors learning and defining a linearity-aware distance to discover the linear manifold space of the original data. We simultaneously utilize the generated representation of the linear structure as input for self-expressiveness to pursue an ideal similarity matrix, which establishes an essential connection with the linearization assumption of the self-expressive strategy. Furthermore, we theoretically demonstrate that our measure can accurately describe the level of linear correlation between instances. Finally, our tests demonstrate that the proposed MLSC attains competitive clustering results compared to state-of-the-art approaches on benchmark datasets.

Prompt learning has emerged as a valuable technique in enhancing vision-language models (VLMs) such as CLIP for downstream tasks in specific domains. Existing work mainly focuses on designing various learning forms of prompts, neglecting the potential of prompts as effective distillers for learning from larger teacher models. In this paper, we introduce an unsupervised domain prompt distillation framework, which aims to transfer the knowledge of a larger teacher model to a lightweight target model through prompt-driven imitation using unlabeled domain images. Specifically, our framework consists of two distinct stages. In the initial stage, we pre-train a large CLIP teacher model using domain (few-shot) labels. After pretraining, we leverage the unique decoupled-modality characteristics of CLIP by pre-computing and storing the text features as class vectors only once through the teacher text encoder. In the subsequent stage, the stored class vectors are shared across teacher and student image encoders for calculating the predicted logits. Further, we align the logits of both the teacher and student models via KL divergence, encouraging the student image encoder to generate similar probability distributions to the teacher through the learnable prompts. The proposed prompt distillation process eliminates the reliance on labeled data, enabling the algorithm to leverage a vast amount of unlabeled images within the domain. Finally, the well-trained student image encoders and pre-stored text features (class vectors) are utilized for inference. To our best knowledge, we are the first to (1) perform unsupervised domain-specific prompt-driven knowledge distillation for CLIP, and (2) establish a practical pre-storing mechanism of text features as shared class vectors between teacher and student. Extensive experiments on 11 datasets demonstrate the effectiveness of our method. Code is publicly available at https://github.com/zhengli97/PromptKD.

Similarity-based representation learning has shown impressive capabilities in both supervised (e.g., metric learning) and unsupervised (e.g., contrastive learning) scenarios. Existing approaches effectively constrained the representation difference (i.e., the disagreement between the embeddings of two instances) to fit the corresponding (pseudo) similarity supervision. However, most of them can hardly restrict the variation of representation difference, sometimes leading to overfitting results where the clusters are disordered by drastically changed differences. In this paper, we thus propose a novel difference alignment regularization (DAR) to encourage all representation differences between inter-class instances to be as close as possible, so that the learning algorithm can produce consistent differences to distinguish data points from each other. To this end, we construct a new cross-total-variation (CTV) norm to measure the divergence among representation differences, and we convert it into an equivalent stochastic form for easy optimization. Then, we integrate the proposed regularizer into the empirical loss for difference-aligned similarity learning (DASL), shrinking the hypothesis space and alleviating overfitting. Theoretically, we prove that our regularizer tightens the error bound of the traditional similarity learning. Experiments on multi-domain data demonstrate the superiority of DASL over existing approaches in both supervised metric learning and unsupervised contrastive learning tasks.

In this paper, we study an object synthesis task that combines an object text with an object image to create a new object image. However, most diffusion models struggle with this task, \textit{i.e.}, often generating an object that predominantly reflects either the text or the image due to an imbalance between their inputs. To address this issue, we propose a simple yet effective method called Adaptive Text-Image Harmony (ATIH) to generate novel and surprising objects. First, we introduce a scale factor and an injection step to balance text and image features in cross-attention and to preserve image information in self-attention during the text-image inversion diffusion process, respectively. Second, to better integrate object text and image, we design a balanced loss function with a noise parameter, ensuring both optimal editability and fidelity of the object image. Third, to adaptively adjust these parameters, we present a novel similarity score function that not only maximizes the similarities between the generated object image and the input text/image but also balances these similarities to harmonize text and image integration. Extensive experiments demonstrate the effectiveness of our approach, showcasing remarkable object creations such as colobus-glass jar. https://xzr52.github.io/ATIH/

Subtle variations are invisible to the naked eyes in human physiological signals can reflect important biological and health indicators. Although numerous computer vision methods have been proposed to recover and magnify these changes, most of them either only focus on identifying and recognizing explicit features such as shapes and textures, or are weak in long-term temporal modeling and spatiotemporal interactive perception of implicit biometrics. Therefore, it is difficult for them to robustly overcome various disturbances that affect detection performance. To address these issues, this paper presents TranPhys, a novel remote photoplethysmography (rPPG) network for facial video-based heart rate estimation. Specifically, first, we argue that facial subregions vary over time due to their biological personalities. So we split the input face video into multiple spatiotemporal tubes, build the 3D vision transformer with encoders and decoders to adequately model the high-dimensional representations of the respective regulars in each subregion, and globally coordinate their feedback on the cardiac pulsing waveform. Second, we design the temporal pooling attention to more finely mine the subtle changes hidden in the skin color over time and their long-term contextual rhythm cues. Third, we leverage the self-supervised masked autoencoding paradigm to overcome redundancy to enhance the robustness of our model, and construct the targeted spatiotemporal sampling maps instead of raw input sequences as the pretrained constraint labels to fully inspire self-supervision. We train, validate, and practice our TranPhys on multiple public datasets to demonstrate that our method achieves the competitive performance in remote heart rate estimation.

Contrastive learning (CL) is a prominent technique for self-supervised representation learning, which aims to contrast semantically similar (i.e., positive) and dissimilar (i.e., negative) pairs of examples under different augmented views. Recently, CL has provided unprecedented potential for learning expressive graph representations without external supervision. In graph CL, the negative nodes are typically uniformly sampled from augmented views to formulate the contrastive objective. However, this uniform negative sampling strategy limits the expressive power of contrastive models. To be specific, not all the negative nodes can provide sufficiently meaningful knowledge for effective contrastive representation learning. In addition, the negative nodes that are semantically similar to the anchor are undesirably repelled from it, leading to degraded model performance. To address these limitations, in this article, we devise an adaptive sampling strategy termed "AdaS. " The proposed AdaS framework can be trained to adaptively encode the importance of different negative nodes, so as to encourage learning from the most informative graph nodes. Meanwhile, an auxiliary polarization regularizer is proposed to suppress the adverse impacts of the false negatives and enhance the discrimination ability of AdaS. The experimental results on a variety of real-world datasets firmly verify the effectiveness of our AdaS in improving the performance of graph CL.

Knowledge distillation aims to learn a lightweight student network from a pre-trained teacher network. In practice, existing knowledge distillation methods are usually infeasible when the original training data is unavailable due to some privacy issues and data management considerations. Therefore, data-free knowledge distillation approaches proposed to collect training instances from the Internet. However, most of them have ignored the common distribution shift between the instances from original training data and webly collected data, affecting the reliability of the trained student network. To solve this problem, we propose a novel method dubbed "Knowledge Distillation between Different Distributions" (KD3), which consists of three components. Specifically, we first dynamically select useful training instances from the webly collected data according to the combined predictions of teacher network and student network. Subsequently, we align both the weighted features and classifier parameters of the two networks for knowledge memorization. Meanwhile, we also build a new contrastive learning block called MixDistribution to generate perturbed data with a new distribution for instance alignment, so that the student network can further learn a distribution-invariant representation. Intensive experiments on various benchmark datasets demonstrate that our proposed KD3can outperform the state-of-the-art data-free knowledge distillation approaches.

In this paper, we propose a novel method for single-view 3D style transfer that generates a unique 3D object with both shape and texture transfer. Our focus lies primarily on birds, a popular subject in 3D reconstruction, for which no existing single-view 3D transfer methods have been developed. The method we propose seeks to generate a 3D mesh shape and texture of a bird from two single-view images. To achieve this, we introduce a novel shape transfer generator that comprises a dual residual gated network (DRGNet), and a multi-layer perceptron (MLP). DRGNet extracts the features of source and target images using a shared coordinate gate unit, while the MLP generates spatial coordinates for building a 3D mesh. We also introduce a semantic UV texture transfer module that implements textural style transfer using semantic UV segmentation, which ensures consistency in the semantic meaning of the transferred regions. This module can be widely adapted to many existing approaches. Finally, our method constructs a novel 3D bird using a differentiable renderer. Experimental results on the CUB dataset verify that our method achieves state-of-the-art performance on the single-view 3D style transfer task. Code is available at https://github.com/wrk226/creative_birds.

Standard MSE or MAE loss function is commonly used in limited field-of-vision depth completion, treating each pixel equally under a basic assumption that all pixels have same contribution during optimization. Recently, with the rapid rise of panoramic photography, panoramic depth completion (PDC) has raised increasing attention in 3D computer vision. However, the assumption is inapplicable to panoramic data due to its latitude-wise distortion and high uncertainty nearby textures and edges. To handle these challenges, we propose distortion and uncertainty aware loss (DUL) that consists of a distortion-aware loss and an uncertainty-aware loss. The distortion-aware loss is designed to tackle the panoramic distortion caused by equirectangular projection, whose coordinate transformation relation is used to adaptively calculate the weight of the latitude-wise distortion, distributing uneven importance instead of the equal treatment for each pixel. The uncertainty-aware loss is presented to handle the inaccuracy in non-smooth regions. Specifically, we characterize uncertainty into PDC solutions under Bayesian deep learning framework, where a novel consistent uncertainty estimation constraint is designed to learn the consistency between multiple uncertainty maps of a single panorama. This consistency constraint allows model to produce more precise uncertainty estimation that is robust to feature deformation. Extensive experiments show the superiority of our method over standard loss functions, reaching the state of the art.

Recently, numerous studies have demonstrated the effectiveness of contrastive learning (CL), which learns feature representations by pulling in positive samples while pushing away negative samples. Many successes of CL lie in that there exists semantic consistency between data augmentations of the same instance. In multi-view scenarios, however, CL might cause representation degeneration when the collected multiple views inherently have inconsistent semantic information or their representations subsequently do not capture sufficient discriminative information. To address this issue, we propose a novel framework called SEM: SElf-weighted Multi-view contrastive learning with reconstruction regularization. Specifically, SEM is a general framework where we propose to first measure the discrepancy between pairwise representations and then minimize the corresponding self-weighted contrastive loss, and thus making SEM adaptively strengthen the useful pairwise views and also weaken the unreliable pairwise views. Meanwhile, we impose a self-supervised reconstruction term to regularize the hidden features of encoders, to assist CL in accessing sufficient discriminative information of data. Experiments on public multi-view datasets verified that SEM can mitigate representation degeneration in existing CL methods and help them achieve significant performance improvements. Ablation studies also demonstrated the effectiveness of SEM with different options of weighting strategies and reconstruction terms.

Obtaining a good similarity matrix is extremely important in subspace clustering. Current state-of-the-art methods learn the similarity matrix through self-expressive strategy. However, these methods directly adopt original samples as a set of basis to represent itself linearly. It is difficult to accurately describe the linear relation between samples in the real-world applications, and thus is hard to find an ideal similarity matrix. To better represent the linear relation of samples, we present a subspace clustering model, Linearity-Aware Subspace Clustering (LASC), which can consciously learn the similarity matrix by employing a linearity-aware metric. This is a new subspace clustering method that combines metric learning and subspace clustering into a joint learning framework. In our model, we first utilize the self-expressive strategy to obtain an initial subspace structure and discover a low-dimensional representation of the original data. Subsequently, we use the proposed metric to learn an intrinsic similarity matrix with linearity-aware on the obtained subspace. Based on such a learned similarity matrix, the inter-cluster distance becomes larger than the intra-cluster distances, and thus successfully obtaining a good subspace cluster result. In addition, to enrich the similarity matrix with more consistent knowledge, we adopt a collaborative learning strategy for self-expressive subspace learning and linearity-aware subspace learning. Moreover, we provide detailed mathematical analysis to show that the metric can properly characterize the linear correlation between samples.

We propose a novel style transfer method to quickly create a new visual product with a nice appearance for industrial designers reference. Given a source product, a target product, and an art style image, our method produces a neural warping field that warps the source shape to imitate the geometric style of the target and a neural texture transformation network that transfers the artistic style to the warped source product. Our model, Industrial Style Transfer (InST), consists of large-scale geometric warping (LGW) and interest-consistency texture transfer (ICTT). LGW aims to explore an unsupervised transformation between the shape masks of the source and target products for fitting large-scale shape warping. Furthermore, we introduce a mask smoothness regularization term to prevent the abrupt changes of the details of the source product. ICTT introduces an interest regularization term to maintain important contents of the warped product when it is stylized by using the art style image. Extensive experimental results demonstrate that InST achieves state-of-the-art performance on multiple visual product design tasks, e.g., companies' snail logos and classical bottles (please see Fig. 1). To the best of our knowledge, we are the first to extend the neural style transfer method to create industrial product appearances. Code is available at https://jcyang98.github.io/InST/home.html

Contrastive learning (CL) pretrains feature embeddings to scatter instances in the feature space so that the training data can be well discriminated. Most existing CL techniques usually encourage learning such feature embeddings in the highdimensional space to maximize the instance discrimination. However, this practice may lead to undesired results where the scattering instances are sparsely distributed in the high-dimensional feature space, making it difficult to capture the underlying similarity between pairwise instances. To this end, we propose a novel framework called contrastive learning with low-dimensional reconstruction (CLLR), which adopts a regularized projection layer to reduce the dimensionality of the feature embedding. In CLLR, we build the sparse / low-rank regularizer to adaptively reconstruct a low-dimensional projection space while preserving the basic objective for instance discrimination, and thus successfully learning contrastive embeddings that alleviate the above issue. Theoretically, we prove a tighter error bound for CLLR; empirically, the superiority of CLLR is demonstrated across multiple domains. Both theoretical and experimental results emphasize the significance of learning low-dimensional contrastive embeddings.

Block-diagonal representation (BDR) is an effective subspace clustering method. The existing BDR methods usually obtain a self-expression coefficient matrix from the original features by a shallow linear model. However, the underlying structure of real-world data is often nonlinear, thus those methods cannot faithfully reflect the intrinsic relationship among samples. To address this problem, we propose a novel latent BDR (LBDR) model to perform the subspace clustering on a nonlinear structure, which jointly learns an autoencoder and a BDR matrix. The autoencoder, which consists of a nonlinear encoder and a linear decoder, plays an important role to learn features from the nonlinear samples. Meanwhile, the learned features are used as a new dictionary for a linear model with block-diagonal regularization, which can ensure good performances for spectral clustering. Moreover, we theoretically prove that the learned features are located in the linear space, thus ensuring the effectiveness of the linear model using self-expression. Extensive experiments on various real-world datasets verify the superiority of our LBDR over the state-of-the-art subspace clustering approaches.

Feature selection aims to select strongly relevant features and discard the rest. Recently, embedded feature selection methods, which incorporate feature weights learning into the training process of a classifier, have attracted much attention. However, traditional embedded methods merely focus on the combinatorial optimality of all selected features. They sometimes select the weakly relevant features with satisfactory combination abilities and leave out some strongly relevant features, thereby degrading the generalization performance. To address this issue, we propose a novel embedded framework for feature selection, termed feature selection boosted by unselected features (FSBUF). Specifically, we introduce an extra classifier for unselected features into the traditional embedded model and jointly learn the feature weights to maximize the classification loss of unselected features. As a result, the extra classifier recycles the unselected strongly relevant features to replace the weakly relevant features in the selected feature subset. Our final objective can be formulated as a minimax optimization problem, and we design an effective gradient-based algorithm to solve it. Furthermore, we theoretically prove that the proposed FSBUF is able to improve the generalization ability of traditional embedded feature selection methods. Extensive experiments on synthetic and real-world data sets exhibit the comprehensibility and superior performance of FSBUF.

The deep learning approach to detecting malicious software (malware) is promising but has yet to tackle the problem of dataset shift, namely that the joint distribution of examples and their labels associated with the test set is different from that of the training set. This problem causes the degradation of deep learning models without users’ notice. In order to alleviate the problem, one approach is to let a classifier not only predict the label on a given example but also present its uncertainty (or confidence) on the predicted label, whereby a defender can decide whether to use the predicted label or not. While intuitive and clearly important, the capabilities and limitations of this approach have not been well understood. In this paper, we conduct an empirical study to evaluate the quality of predictive uncertainties of malware detectors. Specifically, we re-design and build 24 Android malware detectors (by transforming four off-the-shelf detectors with six calibration methods) and quantify their uncertainties with nine metrics, including three metrics dealing with data imbalance. Our main findings are: (i) predictive uncertainty indeed helps achieve reliable malware detection in the presence of dataset shift, but cannot cope with adversarial evasion attacks; (ii) approximate Bayesian methods are promising to calibrate and generalize malware detectors to deal with dataset shift, but cannot cope with adversarial evasion attacks; (iii) adversarial evasion attacks can render calibration methods useless, and it is an open problem to quantify the uncertainty associated with the predicted labels of adversarial examples (i.e., it is not effective to use predictive uncertainty to detect adversarial examples).

Generalized zero-shot learning (GZSL) aims to recognize objects from both seen and unseen classes, when only the labeled examples from seen classes are provided. Recent feature generation methods learn a generative model that can synthesize the missing visual features of unseen classes to mitigate the data-imbalance problem in GZSL. However, the original visual feature space is suboptimal for GZSL classification since it lacks discriminative information. To tackle this issue, we propose to integrate the generation model with the embedding model, yielding a hybrid GZSL framework. The hybrid GZSL approach maps both the real and the synthetic samples produced by the generation model into an embedding space, where we perform the final GZSL classification. Specifically, we propose a contrastive embedding (CE) for our hybrid GZSL framework. The proposed contrastive embedding can leverage not only the class-wise supervision but also the instance-wise supervision, where the latter is usually neglected by existing GZSL researches. We evaluate our proposed hybrid GZSL framework with contrastive embedding, named CE-GZSL, on five benchmark datasets. The results show that our CEGZSL method can outperform the state-of-the-arts by a significant margin on three datasets. Our codes are available on https://github.com/Hanzy1996/CE-GZSL.

\emph{Contrastive learning} (CL) pretrains models in a pairwise manner, where given a data point, other data points are all regarded as dissimilar, including some that are \emph{semantically} similar. The issue has been addressed by properly weighting similar and dissimilar pairs as in \emph{positive-unlabeled learning}, so that the objective of CL is \emph{unbiased} and CL is \emph{consistent}. However, in this paper, we argue that this great solution is still not enough: its weighted objective \emph{hides} the issue where the semantically similar pairs are still pushed away; as CL is pretraining, this phenomenon is not our desideratum and might affect downstream tasks. To this end, we propose \emph{large-margin contrastive learning} (LMCL) with \emph{distance polarization regularizer}, motivated by the distribution characteristic of pairwise distances in \emph{metric learning}. In LMCL, we can distinguish between \emph{intra-cluster} and \emph{inter-cluster} pairs, and then only push away inter-cluster pairs, which \emph{solves} the above issue explicitly. Theoretically, we prove a tighter error bound for LMCL; empirically, the superiority of LMCL is demonstrated across multiple domains, \emph{i.e.}, image classification, sentence representation, and reinforcement learning.

Deep-learning based Super-Resolution (SR) methods have exhibited promising performance under non-blind setting where blur kernel is known; however, blur kernels of Low-Resolution (LR) images in different practical applications are usually unknown. It may lead to a significant performance drop when degradation process of training images deviates from that of real images. In this paper, we propose a novel blind SR framework to super-resolve LR images degraded by arbitrary blur kernel with accurate kernel estimation in frequency domain. To our best knowledge, this is the first deep learning method which conducts blur kernel estimation in frequency domain. Specifically, we first demonstrate that feature representation in frequency domain is more conducive for blur kernel reconstruction than in spatial domain. Next, we present a Spectrum-to-Kernel (S$2$K) network to estimate general blur kernels in diverse forms. We use a conditional GAN (CGAN) combined with SR-oriented optimization target to learn the end-to-end translation from degraded images' spectra to unknown kernels. Extensive experiments on both synthetic and real-world images demonstrate that our proposed method sufficiently reduces blur kernel estimation error, thus enables the off-the-shelf non-blind SR methods to work under blind setting effectively, and achieves superior performance over state-of-the-art blind SR methods, averagely by 1.39dB, 0.48dB (Gaussian kernels) and 6.15dB, 4.57dB (motion kernels) for scales $2\times$ and $4\times$ respectively.

Inspired by biological evolution, we explain the rationality of Vision Transformer by analogy with the proven practical Evolutionary Algorithm (EA) and derive that both of them have consistent mathematical representation. Analogous to the dynamic local population in EA, we improve the existing transformer structure and propose a more efficient EAT model, and design task-related heads to deal with different tasks more flexibly. Moreover, we introduce the spatial-filling curve into the current vision transformer to sequence image data into a uniform sequential format. Thus we can design a unified EAT framework to address multi-modal tasks, separating the network architecture from the data format adaptation. Our approach achieves state-of-the-art results on the ImageNet classification task compared with recent vision transformer works while having smaller parameters and greater throughput. We further conduct multi-modal tasks to demonstrate the superiority of the unified EAT, \eg, Text-Based Image Retrieval, and our approach improves the rank-1 by +3.7 points over the baseline on the CSS dataset.

-norm, etc.) have been widely leveraged to form the loss function of different regression models, and have played an important role in image analysis. However, the previous regression models adopting the existing norms are sensitive to outliers and, thus, often bring about unsatisfactory results on the heavily corrupted images. This is because their adopted norms for measuring the data residual can hardly suppress the negative influence of noisy data, which will probably mislead the regression process. To address this issue, this paper proposes a novel δ (delta)-norm to count the nonzero blocks around an element in a vector or matrix, which weakens the impacts of outliers and also takes the structure property of examples into account. After that, we present the δ -norm-based robust regression (DRR) in which the data examples are mapped to the kernel space and measured by the proposed δ -norm. By exploring an explicit kernel function, we show that DRR has a closed-form solution, which suggests that DRR can be efficiently solved. To further handle the influences from mixed noise, DRR is extended to a multiscale version. The experimental results on image classification and background modeling datasets validate the superiority of the proposed approach to the existing state-of-the-art robust regression models.

The merits of fast convergence and potentially better performance of the weight normalization family have drawn increasing attention in recent years. These methods use standardization or normalization that changes the weight W to W′, which makes W′ independent to the magnitude of W. Surprisingly, W must be decayed during gradient descent, otherwise we will observe a severe under-fitting problem, which is very counter-intuitive since weight decay is widely known to prevent deep networks from over-fitting. Moreover, if we substitute (e.g., weight normalization) W′ = W∥W∥ in the original loss function ∑i L(ƒ(xi; W′),yi) + ½λ∥W′∥2, it is observed that the regularization term ½λ∥W′∥2 will be canceled as a constant ½ λ in the optimization objective. Therefore, to decay W, we need to explicitly append: ½λ∥W∥2. In this paper, we theoretically prove that ½λ∥W∥2 improves optimization only by modulating the effective learning rate and fairly has no influence on generalization when the weight normalization family is compositely employed. Furthermore, we also expose several serious problems when introducing weight decay term to weight normalization family, including the missing of global minimum, training instability and sensitivity of initialization. To address these problems, we propose an Adaptive Weight Shrink (AWS) scheme, which gradually shrinks the weights during optimization by a dynamic coefficient proportional to the magnitude of the parameter. This simple yet effective method appropriately controls the effective learning rate, which significantly improves the training stability and makes optimization more robust to initialization.

In human multi-modality perception systems, the benefits of integrating auditory and visual information are extensive as they provide plenty supplementary cues for understanding the events. Despite some recent methods proposed for such application, they cannot deal with practical conditions with temporal inconsistency. Inspired by human system which puts different focuses at specific locations, time segments and media while performing multi-modality perception, we provide an attention-based method to simulate such process. Similar to human mechanism, our network can adaptively select “where” to attend, “when” to attend and “which” to attend for audio-visual event localization. In this way, even with large temporal inconsistent between vision and audio, our network is able to adaptively trade information between different modalities and successfully achieve event localization. Our method achieves state-of-the-art performance on AVE (Audio-Visual Event) dataset collected in the real life. In addition, we also systemically investigate audio-visual event localization tasks. The visualization results also help us better understand how our model works.

One-stage detector basically formulates object detection as dense classification and localization (i.e., bounding box regression). The classification is usually optimized by Focal Loss and the box location is commonly learned under Dirac delta distribution. A recent trend for one-stage detectors is to introduce an \emph{individual} prediction branch to estimate the quality of localization, where the predicted quality facilitates the classification to improve detection performance. This paper delves into the \emph{representations} of the above three fundamental elements: quality estimation, classification and localization. Two problems are discovered in existing practices, including (1) the inconsistent usage of the quality estimation and classification between training and inference, and (2) the inflexible Dirac delta distribution for localization. To address the problems, we design new representations for these elements. Specifically, we merge the quality estimation into the class prediction vector to form a joint representation, and use a vector to represent arbitrary distribution of box locations. The improved representations eliminate the inconsistency risk and accurately depict the flexible distribution in real data, but contain \emph{continuous} labels, which is beyond the scope of Focal Loss. We then propose Generalized Focal Loss (GFL) that generalizes Focal Loss from its discrete form to the \emph{continuous} version for successful optimization. On COCO {\tt test-dev}, GFL achieves 45.0\% AP using ResNet-101 backbone, surpassing state-of-the-art SAPD (43.5\%) and ATSS (43.6\%) with higher or comparable inference speed.

Practical data sets often contain the label noise caused by various human factors or measurement errors, which means that a fraction of training examples might be mistakenly labeled. Such noisy labels will mislead the classifier training and severely decrease the classification performance. Existing approaches to handle this problem are usually developed through various surrogate loss functions under the framework of empirical risk minimization. However, they are only suitable for binary classification and also require strong prior knowledge. Therefore, this article treats the example features as side information and formulates the noisy label removal problem as a matrix recovery problem. We denote our proposed method as "label noise handling via side information" (LNSI). Specifically, the observed label matrix is decomposed as the sum of two parts, in which the first part reveals the true labels and can be obtained by conducting a low-rank mapping on the side information; and the second part captures the incorrect labels and is modeled by a row-sparse matrix. The merits of such formulation lie in three aspects: 1) the strong recovery ability of this strategy has been sufficiently demonstrated by intensive theoretical works on side information; 2) multi-class situations can be directly handled with the aid of learned projection matrix; and 3) only very weak assumptions are required for model design, making LNSI applicable to a wide range of practical problems. Moreover, we theoretically derive the generalization bound of LNSI and show that the expected classification error of LNSI is upper bounded. The experimental results on a variety of data sets including UCI benchmark data sets and practical data sets confirm the superiority of LNSI to state-of-the-art approaches on label noise handling.

The central problem for most existing metric learning methods is to find a suitable projection matrix on the differences of all pairs of data points. However, a single unified projection matrix can hardly characterize all data similarities accurately as the practical data are usually very complicated, and simply adopting one global projection matrix might ignore important local patterns hidden in the dataset. To address this issue, this paper proposes a novel method dubbed “Data-Adaptive Metric Learning” (DAML), which constructs a data-adaptive projection matrix for each data pair by selectively combining a set of learned candidate matrices. As a result, every data pair can obtain a specific projection matrix, enabling the proposed DAML to flexibly fit the training data and produce discriminative projection results. The model of DAML is formulated as an optimization problem which jointly learns candidate projection matrices and their sparse combination for every data pair. Nevertheless, the over-fitting problem may occur due to the large amount of parameters to be learned. To tackle this issue, we adopt the Total Variation (TV) regularizer to align the scales of data embedding produced by all candidate projection matrices, and thus the generated metrics of these learned candidates are generally comparable. Furthermore, we extend the basic linear DAML model to the kernerlized version (denoted “KDAML”) to handle the non-linear cases, and the Iterative Shrinkage-Thresholding Algorithm (ISTA) is employed to solve the optimization model. Intensive experimental results on various applications including retrieval, classification, and verification clearly demonstrate the superiority of our algorithm to other state-of-the-art metric learning methodologies.

This paper first answers the question ``why do the two most powerful techniques Dropout and Batch Normalization (BN) often lead to a worse performance when they are combined together in many modern neural networks, but cooperate well sometimes as in Wide ResNet (WRN)?'' in both theoretical and empirical aspects. Theoretically, we find that Dropout shifts the variance of a specific neural unit when we transfer the state of that network from training to test. However, BN maintains its statistical variance, which is accumulated from the entire learning procedure, in the test phase. The inconsistency of variances in Dropout and BN (we name this scheme ``variance shift'') causes the unstable numerical behavior in inference that leads to erroneous predictions finally. Meanwhile, the large feature dimension in WRN further reduces the ``variance shift'' to bring benefits to the overall performance. Thorough experiments on representative modern convolutional networks like DenseNet, ResNet, ResNeXt and Wide ResNet confirm our findings. According to the uncovered mechanism, we get better understandings in the combination of these two techniques and summarize guidelines for better practices.

Distance Metric Learning aims to learn an appropriate metric that faithfully measures the distance between two data points. Traditional metric learning methods usually calculate the pairwise distance with fixed distance functions (\emph{e.g.,}\ Euclidean distance) in the projected feature spaces. However, they fail to learn the underlying geometries of the sample space, and thus cannot exactly predict the intrinsic distances between data points. To address this issue, we first reveal that the traditional linear distance metric is equivalent to the cumulative arc length between the data pair's nearest points on the learned straight measurer lines. After that, by extending such straight lines to general curved forms, we propose a Curvilinear Distance Metric Learning (CDML) method, which adaptively learns the nonlinear geometries of the training data. By virtue of Weierstrass theorem, the proposed CDML is equivalently parameterized with a 3-order tensor, and the optimization algorithm is designed to learn the tensor parameter. Theoretical analysis is derived to guarantee the effectiveness and soundness of CDML. Extensive experiments on the synthetic and real-world datasets validate the superiority of our method over the state-of-the-art metric learning models.

In the past decades, intensive efforts have been put to design various loss functions and metric forms for metric learning problem. These improvements have shown promising results when the test data is similar to the training data. However, the trained models often fail to produce reliable distances on the ambiguous test pairs due to the different samplings between training set and test set. To address this problem, the Adversarial Metric Learning (AML) is proposed in this paper, which automatically generates adversarial pairs to remedy the sampling bias and facilitate robust metric learning. Specifically, AML consists of two adversarial stages, i.e. confusion and distinguishment. In confusion stage, the ambiguous but critical adversarial data pairs are adaptively generated to mislead the learned metric. In distinguishment stage, a metric is exhaustively learned to try its best to distinguish both adversarial pairs and original training pairs. Thanks to the challenges posed by the confusion stage in such competing process, the AML model is able to grasp plentiful difficult knowledge that has not been contained by the original training pairs, so the discriminability of AML can be significantly improved. The entire model is formulated into optimization framework, of which the global convergence is theoretically proved. The experimental results on toy data and practical datasets clearly demonstrate the superiority of AML to representative state-of-the-art metric learning models.

This paper develops a novel method to address the structural noise in samples for image classification. Recently, regression-related classification methods have shown promising results when facing the pixelwise noise. However, they become weak in coping with the structural noise due to ignoring of relationships between pixels of noise image. Meanwhile, most of them need to implement the iterative process for computing representation coefficients, which leads to the high time consumption. To overcome these problems, we exploit a latent pattern model called low-rank latent pattern approximation (LLPA) to reconstruct the test image having structural noise. The rank function is applied to characterize the structure of the reconstruction residual between test image and the corresponding latent pattern. Simultaneously, the error between the latent pattern and the reference image is constrained by Frobenius norm to prevent overfitting. LLPA involves a closed-form solution by the virtue of a singular value thresholding operator. The provided theoretic analysis demonstrates that LLPA indeed removes the structural noise during classification task. Additionally, LLPA is further extended to the form of matrix regression by connecting multiple training samples, and alternating direction of multipliers method with Gaussian back substitution algorithm is used to solve the extended LLPA. Experimental results on several popular data sets validate that the proposed methods are more robust to image classification with occlusion and illumination changes, as compared to some existing state-of-the-art reconstruction-based methods and one deep neural network-based method.

Influence maximization, whose objective is to select k users (called seeds) from a social network such that the number of users influenced by the seeds (called influence spread) is maximized, has attracted significant attention due to its widespread applications, such as viral marketing and rumor control. However, in real-world social networks, users have their own interests (which can be represented as topics) and are more likely to be influenced by their friends (or friends' friends) with similar topics. We can increase the influence spread by taking into consideration topics. To address this problem, we study topic-aware influence maximization, which, given a topic-aware influence maximization (TIM) query, finds k seeds from a social network such that the topic-aware influence spread of the k seeds is maximized. Our goal is to enable online TIM queries. Since the topic-aware influence maximization problem is NP-hard, we focus on devising efficient algorithms to achieve instant performance while keeping a high influence spread. We utilize a maximum influence arborescence (MIA) model to approximate the computation of influence spread. To efficiently find k seeds under the MIA model, we first propose a best-effort algorithm with 1 − 1/ e approximation ratio, which estimates an upper bound of the topic-aware influence of each user and utilizes the bound to prune large numbers of users with small influence. We devise effective techniques to estimate tighter upper bounds. We then propose a faster topic-sample-based algorithm with ε · (1 − 1/ e ) approximation ratio for any ε ∈ (0, 1], which materializes the influence spread of some topic-distribution samples and utilizes the materialized information to avoid computing the actual influence of users with small influences. Experimental results show that our methods significantly outperform baseline approaches.

users in a social network to maximize the expected number of users influenced by the selected users (called influence spread), has been extensively studied, existing works neglected the fact that the location information can play an important role in influence maximization. Many real-world applications such as location-aware word-of-mouth marketing have location-aware requirement. In this paper we study the location-aware influence maximization problem. One big challenge in location-aware influence maximization is to develop an efficient scheme that offers wide influence spread. To address this challenge, we propose two greedy algorithms with 1-1/e approximation ratio. To meet the instant-speed requirement, we propose two efficient algorithms with ε· (1-1/e) approximation ratio for any ε ∈ (0,1]. Experimental results on real datasets show our method achieves high performance while keeping large influence spread and significantly outperforms state-of-the-art algorithms.

Existing route recommendation systems have two main weaknesses. First, they usually recommend the same route for all users and cannot help control traffic jam. Second, they do not take full advantage of real-time traffic to recommend the best routes. To address these two problems, we develop a real-time route recommendation system, called R3, aiming to provide users with the real-time-traffic-aware routes. R3 recommends diverse routes for different users to alleviate the traffic pressure. R3 utilizes historical taxi driving data and real-time traffic data and integrates them together to provide users with real-time route recommendation.