Mixture-of-Experts (MoE) is widely adopted to deploy Large Language Models (LLMs) on edge devices with limited memory budgets. Although MoE is, in theory, an inborn memory-friendly architecture requiring only a few activated experts to reside in the memory for inference, current MoE architectures cannot effectively fulfill this advantage and will yield intolerable inference latencies of LLMs on memory-constrained devices. Our investigation pinpoints the essential cause as the remarkable temporal inconsistencies of inter-token expert activations, which generate overly frequent expert swapping demands dominating the latencies. To this end, we propose a novel MoE architecture, Oracle-MoE, to fulfill the real on-device potential of MoE-based LLMs. Oracle-MoE route tokens in a highly compact space suggested by attention scores, termed the oracle space, to effectively maintain the semantic locality across consecutive tokens to reduce expert activation variations, eliminating massive swapping demands. Theoretical analysis proves that Oracle-MoE is bound to provide routing decisions with better semantic locality and, therefore, better expert activation consistencies. Experiments on the pretrained GPT-2 architectures of different sizes (200M, 350M, 790M, and 2B) and downstream tasks demonstrate that without compromising task performance, our Oracle-MoE has achieved state-of-the-art inference speeds across varying memory budgets, revealing its substantial potential for LLM deployments in industry.

Denoising-based diffusion models have attained impressive image synthesis; however, their applications on videos can lead to unaffordable computational costs due to the per-frame denoising operations. In pursuit of efficient video generation, we present a Diffusion Reuse MOtion (Dr. Mo) network to accelerate the video-based denoising process. Our crucial observation is that the latent representations in early denoising steps between adjacent video frames exhibit high consistencies with motion clues. Inspired by the discovery, we propose to accelerate the video denoising process by incorporating lightweight, learnable motion features. Specifically, Dr. Mo will only compute all denoising steps for base frames. For a non-based frame, Dr. Mo will propagate the pre-computed based latents of a particular step with inter-frame motions to obtain a fast estimation of its coarse-grained latent representation, from which the denoising will continue to obtain more sensitive and fine-grained representations. On top of this, Dr. Mo employs a meta-network named Denoising Step Selector (DSS) to dynamically determine the step to perform motion-based propagations for each frame, ensuring the correct transformation of multi-granularity visual features. Extensive evaluations on video generation and editing tasks indicate that Dr. Mo delivers widely applicable acceleration for diffusion-based video generations while effectively retaining the visual quality and style. Video generation and visualization results can be found athttps://drmo-denoising-reuse.github.io.

Language models (LMs) only pretrained on a general and massive corpus usually cannot attain satisfying performance on domain-specific downstream tasks, and hence, applying domain-specific pretraining to LMs is a common and indispensable practice. However, domain-specific pretraining can be costly and time-consuming, hindering LMs' deployment in real-world applications. In this work, we consider the incapability to memorize domain-specific knowledge embedded in the general corpus with rare occurrences and long-tail distributions as the leading cause for pretrained LMs' inferior downstream performance. Analysis of Neural Tangent Kernels (NTKs) reveals that those long-tail data are commonly overlooked in the model's gradient updates and, consequently, are not effectively memorized, leading to poor domain-specific downstream performance. Based on the intuition that data with similar semantic meaning are closer in the embedding space, we devise a Cluster-guided Sparse Expert (CSE) layer to actively learn long-tail domain knowledge typically neglected in previous pretrained LMs. During pretraining, a CSE layer efficiently clusters domain knowledge together and assigns long-tail knowledge to designate extra experts. CSE is also a lightweight structure that only needs to be incorporated in several deep layers. With our training strategy, we found that during pretraining, data of long-tail knowledge gradually formulate isolated, outlier clusters in an LM's representation spaces, especially in deeper layers. Our experimental results show that only pretraining CSE-based LMs is enough to achieve superior performance than regularly pretrained-finetuned LMs on various downstream tasks, implying the prospects of domain-specific-pretraining-free language models.

Metric learning aims to learn a distance metric such that semantically similar instances are pulled together while dissimilar instances are pushed away. Many existing methods consider maximizing or at least constraining a distance margin in the feature space that separates similar and dissimilar pairs of instances to guarantee their generalization ability. In this article, we advocate imposing an adversarial margin in the input space so as to improve the generalization and robustness of metric learning algorithms. We first show that the adversarial margin, defined as the distance between training instances and their closest adversarial examples in the input space, takes account of both the distance margin in the feature space and the correlation between the metric and triplet constraints. Next, to enhance robustness to instance perturbation, we propose to enlarge the adversarial margin through minimizing a derived novel loss function termed the perturbation loss. The proposed loss can be viewed as a data-dependent regularizer and easily plugged into any existing metric learning methods. Finally, we show that the enlarged margin is beneficial to the generalization ability by using the theoretical technique of algorithmic robustness. Experimental results on 16 datasets demonstrate the superiority of the proposed method over existing state-of-the-art methods in both discrimination accuracy and robustness against possible noise.

Despite their prevalence in deep-learning communities, over-parameterized models convey high demands of computational costs for proper training. This work studies the fine-grained, modular-level learning dynamics of over-parameterized models to attain a more efficient and fruitful training strategy. Empirical evidence reveals that when scaling down into network modules, such as heads in self-attention models, we can observe varying learning patterns implicitly associated with each module's trainability. To describe such modular-level learning capabilities, we introduce a novel concept dubbed modular neural tangent kernel (mNTK), and we demonstrate that the quality of a module's learning is tightly associated with its mNTK's principal eigenvalue $\lambda_{\max}$. A large $\lambda_{\max}$ indicates that the module learns features with better convergence, while those miniature ones may impact generalization negatively. Inspired by the discovery, we propose a novel training strategy termed Modular Adaptive Training (MAT) to update those modules with their $\lambda_{\max}$ exceeding a dynamic threshold selectively, concentrating the model on learning common features and ignoring those inconsistent ones. Unlike most existing training schemes with a complete BP cycle across all network modules, MAT can significantly save computations by its partially-updating strategy and can further improve performance. Experiments show that MAT nearly halves the computational cost of model training and outperforms the accuracy of baselines.

This article presents a novel method for face clustering in videos using a video-centralised transformer. Previous works often employed contrastive learning to learn frame-level representation and used average pooling to aggregate the features along the temporal dimension. This approach may not fully capture the complicated video dynamics. In addition, despite the recent progress in video-based contrastive learning, few have attempted to learn a self-supervised clustering-friendly face representation that benefits the video face clustering task. To overcome these limitations, our method employs a transformer to directly learn video-level representations that can better reflect the temporally-varying property of faces in videos, while we also propose a video-centralised self-supervised framework to train the transformer model. We also investigate face clustering in egocentric videos, a fast-emerging field that has not been studied yet in works related to face clustering. To this end, we present and release the first large-scale egocentric video face clustering dataset named EasyCom-Clustering. We evaluate our proposed method on both the widely used Big Bang Theory (BBT) dataset and the new EasyCom-Clustering dataset. Results show the performance of our video-centralised transformer has surpassed all previous state-of-the-art methods on both benchmarks, exhibiting a self-attentive understanding of face videos.

Chip floorplanning has long been a critical task with high computation complexity in the physical implementation of VLSI chips. Its key objective is to determine the initial locations of large chip modules with minimized wirelength while adhering to the density constraint, which in essence is a process of constructing an optimized mapping from circuit connectivity to physical locations. Proven to be an NP-hard problem, chip floorplanning is difficult to be solved efficiently using algorithmic approaches. This article presents GraphPlanner, a variational graph-convolutional-network-based deep learning technique for chip floorplanning. GraphPlanner is able to learn an optimized and generalized mapping between circuit connectivity and physical wirelength and produce a chip floorplan using efficient model inference. GraphPlanner is further equipped with an efficient clustering method, a unification of hyperedge coarsening with graph spectral clustering, to partition a large-scale netlist into high-quality clusters with minimized inter-cluster weighted connectivity. GraphPlanner has been integrated with two state-of-the-art mixed-size placers. Experimental studies using both academic benchmarks and industrial designs demonstrate that compared to state-of-the-art mixed-size placers alone, GraphPlanner improves placement runtime by 25% with 4% wirelength reduction on average.

Floorplanning has long been a critical physical design task with high computation complexity. Its key objective is to determine the initial locations of macros and standard cells with optimized wirelength for a given area constraint. This paper presents Flora, a graph attention-based floorplanner to learn an optimized mapping between circuit connectivity and physical wirelength, and produce a chip floorplan using efficient model inference. Flora has been integrated with two state-of-the-art mixed-size placers. Experimental studies using both academic benchmarks and industrial designs demonstrate that compared to state-of-the-art mixed-size placers alone, Flora improves placement runtime by 18%, with 2% wirelength reduction on average.

Deep-learning-based video processing has yielded transformative results in recent years. However, the video analytics pipeline is energy-intensive due to high data rates and reliance on complex inference algorithms, which limits its adoption in energy-constrained applications. Motivated by the observation of high and variable spatial redundancy and temporal dynamics in video data streams, we design and evaluate an adaptive-resolution optimization framework to minimize the energy use of multi-task video analytics pipelines. Instead of heuristically tuning the input data resolution of individual tasks, our framework utilizes deep reinforcement learning to dynamically govern the input resolution and computation of the entire video analytics pipeline. By monitoring the impact of varying resolution on the quality of high-dimensional video analytics features, hence the accuracy of video analytics results, the proposed end-to-end optimization framework learns the best non-myopic policy for dynamically controlling the resolution of input video streams to globally optimize energy efficiency. Governed by reinforcement learning, optical flow is incorporated into the framework to minimize unnecessary spatio-temporal redundancy that leads to re-computation, while preserving accuracy. The proposed framework is applied to video instance segmentation which is one of the most challenging computer vision tasks, and achieves better energy efficiency than all baseline methods of similar accuracy on the YouTube-VIS dataset.

For real-time semantic video segmentation, most recent works utilised a dynamic framework with a key scheduler to make online key/non-key decisions. Some works used a fixed key scheduling policy, while others proposed adaptive key scheduling methods based on heuristic strategies, both of which may lead to suboptimal global performance. To overcome this limitation, we model the online key decision process in dynamic video segmentation as a deep reinforcement learning problem and learn an efficient and effective scheduling policy from expert information about decision history and from the process of maximising global return. Moreover, we study the application of dynamic video segmentation on face videos, a field that has not been investigated before. By evaluating on the 300VW dataset, we show that the performance of our reinforcement key scheduler outperforms that of various baselines in terms of both effective key selections and running speed. Further results on the Cityscapes dataset demonstrate that our proposed method can also generalise to other scenarios. To the best of our knowledge, this is the first work to use reinforcement learning for online key-frame decision in dynamic video segmentation, and also the first work on its application on face videos.

Metric learning aims to learn a distance measure that can benefit distance-based methods such as the nearest neighbour (NN) classifier. While considerable efforts have been made to improve its empirical performance and analyze its generalization ability by focusing on the data structure and model complexity, an unresolved question is how choices of algorithmic parameters, such as the number of training iterations, affect metric learning as it is typically formulated as an optimization problem and nowadays more often as a non-convex problem. In this paper, we theoretically address this question and prove the agnostic Probably Approximately Correct (PAC) learnability for metric learning algorithms with non-convex objective functions optimized via gradient descent (GD); in particular, our theoretical guarantee takes the iteration number into account. We first show that the generalization PAC bound is a sufficient condition for agnostic PAC learnability and this bound can be obtained by ensuring the uniform convergence on a densely concentrated subset of the parameter space. We then show that, for classifiers optimized via GD, their generalizability can be guaranteed if the classifier and loss function are both Lipschitz smooth, and further improved by using fewer iterations. To illustrate and exploit the theoretical findings, we finally propose a novel metric learning method called Smooth Metric and representative Instance LEarning (SMILE), designed to satisfy the Lipschitz smoothness property and learned via GD with an early stopping mechanism for better discriminability and less computational cost of NN.

The performance of distance-based classifiers heavily depends on the underlying distance metric, so it is valuable to learn a suitable metric from the data. To address the problem of multimodality, it is desirable to learn local metrics. In this short paper, we define a new intuitive distance with local metrics and influential regions, and subsequently propose a novel local metric learning algorithm called LMLIR for distance-based classification. Our key intuition is to partition the metric space into influential regions and a background region, and then regulate the effectiveness of each local metric to be within the related influential regions. We learn multiple local metrics and influential regions to reduce the empirical hinge loss, and regularize the parameters on the basis of a resultant learning bound. Encouraging experimental results are obtained from various public and popular data sets.

Active learning aims to reduce annotation cost by predicting which samples are useful for a human teacher to label. However it has become clear there is no best active learning algorithm. Inspired by various philosophies about what constitutes a good criteria, different algorithms perform well on different datasets. This has motivated research into ensembles of active learners that learn what constitutes a good criteria in a given scenario, typically via multi-armed bandit algorithms. Though algorithm ensembles can lead to better results, they overlook the fact that not only does algorithm efficacy vary across datasets, but also during a single active learning session. That is, the best criteria is non-stationary. This breaks existing algorithms' guarantees and hampers their performance in practice. In this paper, we propose dynamic ensemble active learning as a more general and promising research direction. We develop a dynamic ensemble active learner based on a non-stationary multi-armed bandit with expert advice algorithm. Our dynamic ensemble selects the right criteria at each step of active learning. It has theoretical guarantees, and shows encouraging results on 13 popular datasets.

Image tagging is a well known challenge in image processing. It is typically addressed through multi-instance multi-label (MIML) classification methodologies. Convolutional Neural Networks (CNNs) possess great potential to perform well on MIML tasks, since multi-level convolution and max pooling coincide with the multi-instance setting and the sharing of hidden representation may benefit multi-label modeling. However, CNNs usually require a large amount of carefully labeled data for training, which is hard to obtain in many real applications. In this paper, we propose a new approach for transferring pre-trained deep networks such as VGG16 on Imagenet to small MIML tasks. We extract features from each group of the network layers and apply multiple binary classifiers to them for multi-label prediction. Moreover, we adopt an L1-norm regularized Logistic Regression (L1LR) to find the most effective features for learning the multi-label classifiers. The experiment results on two most-widely used and relatively small benchmark MIML image datasets demonstrate that the proposed approach can substantially outperform the state-of-the-art algorithms, in terms of all popular performance metrics.

Facial Expression Recognition (FER) techniques have already been adopted in numerous multimedia systems. Plenty of previous research assumes that each facial picture should be linked to only one of the predefined affective labels. Nevertheless, in practical applications, few of the expressions are exactly one of the predefined affective states. Therefore, to depict the facial expressions more accurately, this paper proposes a multi-label classification approach for FER and each facial expression would be labeled with one or multiple affective states. Meanwhile, by modeling the relationship between labels via Group Lasso regularization term, a maximum margin multi-label classifier is presented and the convex optimization formulation guarantees a global optimal solution. To evaluate the performance of our classifier, the JAFFE dataset is extended into a multi-label facial expression dataset by setting threshold to its continuous labels marked in the original dataset and the labeling results have shown that multiple labels can output a far more accurate description of facial expression. At the same time, the classification results have verified the superior performance of our algorithm.

To alleviate the workload of labeling before estimating certain color distributions, integrative labeling is introduced, which merely needs to figure out whether a picture contains positive-class regions or not and then all pixels of the picture are treated as positive or negative class training samples. Integrative labeling, however, results in heavy mixture of training samples. Thus traditional generative density estimation methods can't be used directly in that they perform poorly with heavily polluted training samples. In this paper, by utilizing the prior knowledge of high separability between positive and negative class color distributions, a discriminative learning based GMM(DiscGMM) is proposed for integrative labeling. Besides generating the polluted positive-class samples with comparatively high probability, optimal parameters found by DiscGMM also enjoy a comparatively low probability of generating negative-class samples. The parameter learning problem is solved by a modified Expectation Maximization (EM) algorithm. In an integrative labeling experiment of skin detection, DiscGMM is testified to enjoy much better performance than generative density estimation methods and shows qualified results.