Xiaoli Li

Kaixin Xu, Zhe Wang, Chunyun Chen et al.

Vision transformers have emerged as a promising alternative to convolutional neural networks for various image analysis tasks, offering comparable or superior performance. However, one significant drawback of ViTs is their resource-intensive nature, leading to increased memory footprint, computation complexity, and power consumption. To democratize this high-performance technology and make it more environmentally friendly, it is essential to compress ViT models, reducing their resource requirements while maintaining high performance. In this paper, we introduce a new block-structured pruning to address the resource-intensive issue for ViTs, offering a balanced trade-off between accuracy and hardware acceleration. Unlike unstructured pruning or channel-wise structured pruning, block pruning leverages the block-wise structure of linear layers, resulting in more efficient matrix multiplications. To optimize this pruning scheme, our paper proposes a novel hardware-aware learning objective that simultaneously maximizes speedup and minimizes power consumption during inference, tailored to the block sparsity structure. This objective eliminates the need for empirical look-up tables and focuses solely on reducing parametrized layer connections. Moreover, our paper provides a lightweight algorithm to achieve post-training pruning for ViTs, utilizing second-order Taylor approximation and empirical optimization to solve the proposed hardware-aware objective. Extensive experiments on ImageNet are conducted across various ViT architectures, including DeiT-B and DeiT-S, demonstrating competitive performance with other pruning methods and achieving a remarkable balance between accuracy preservation and power savings. Especially, we achieve 3.93x speedup on dedicated hardware and GPUs respectively for DeiT-B, and a power reduction by 1.4x on GPUs. Code released to https://github.com/Akimoto-Cris/LPViT.

Qing Xu, Min Wu, Xiaoli Li et al.

For many real-world time series tasks, the computational complexity of prevalent deep leaning models often hinders the deployment on resource-limited environments (e.g., smartphones). Moreover, due to the inevitable domain shift between model training (source) and deploying (target) stages, compressing those deep models under cross-domain scenarios becomes more challenging. Although some of existing works have already explored cross-domain knowledge distillation for model compression, they are either biased to source data or heavily tangled between source and target data. To this end, we design a novel end-to-end framework called Universal and joint knowledge distillation (UNI-KD) for cross-domain model compression. In particular, we propose to transfer both the universal feature-level knowledge across source and target domains and the joint logit-level knowledge shared by both domains from the teacher to the student model via an adversarial learning scheme. More specifically, a feature-domain discriminator is employed to align teacher's and student's representations for universal knowledge transfer. A data-domain discriminator is utilized to prioritize the domain-shared samples for joint knowledge transfer. Extensive experimental results on four time series datasets demonstrate the superiority of our proposed method over state-of-the-art (SOTA) benchmarks.

Ruibing Jin, Guosheng Lin, Min Wu et al.

Dark videos often lose essential information, which causes the knowledge learned by networks is not enough to accurately recognize actions. Existing knowledge assembling methods require massive GPU memory to distill the knowledge from multiple teacher models into a student model. In action recognition, this drawback becomes serious due to much computation required by video process. Constrained by limited computation source, these approaches are infeasible. To address this issue, we propose an unlimited knowledge distillation (UKD) in this paper. Compared with existing knowledge assembling methods, our UKD can effectively assemble different knowledge without introducing high GPU memory consumption. Thus, the number of teaching models for distillation is unlimited. With our UKD, the network's learned knowledge can be remarkably enriched. Our experiments show that the single stream network distilled with our UKD even surpasses a two-stream network. Extensive experiments are conducted on the ARID dataset.

Hezhe Qiao, Qingsong Wen, Xiaoli Li et al.

This work considers a practical semi-supervised graph anomaly detection (GAD) scenario, where part of the nodes in a graph are known to be normal, contrasting to the extensively explored unsupervised setting with a fully unlabeled graph. We reveal that having access to the normal nodes, even just a small percentage of normal nodes, helps enhance the detection performance of existing unsupervised GAD methods when they are adapted to the semi-supervised setting. However, their utilization of these normal nodes is limited. In this paper, we propose a novel Generative GAD approach (namely GGAD) for the semi-supervised scenario to better exploit the normal nodes. The key idea is to generate pseudo anomaly nodes, referred to as 'outlier nodes', for providing effective negative node samples in training a discriminative one-class classifier. The main challenge here lies in the lack of ground truth information about real anomaly nodes. To address this challenge, GGAD is designed to leverage two important priors about the anomaly nodes -- asymmetric local affinity and egocentric closeness -- to generate reliable outlier nodes that assimilate anomaly nodes in both graph structure and feature representations. Comprehensive experiments on six real-world GAD datasets are performed to establish a benchmark for semi-supervised GAD and show that GGAD substantially outperforms state-of-the-art unsupervised and semi-supervised GAD methods with varying numbers of training normal nodes. Code will be made available at https://github.com/mala-lab/GGAD.

Kaixin Xu, Qingtian Feng, Hao Chen et al.

Applying deep neural networks to 3D point cloud processing has attracted increasing attention due to its advanced performance in many areas, such as AR/VR, autonomous driving, and robotics. However, as neural network models and 3D point clouds expand in size, it becomes a crucial challenge to reduce the computational and memory overhead to meet latency and energy constraints in real-world applications. Although existing approaches have proposed to reduce both computational cost and memory footprint, most of them only address the spatial redundancy in inputs, i.e. removing the redundancy of background points in 3D data. In this paper, we propose a novel post-training weight pruning scheme for 3D object detection that is (1) orthogonal to all existing point cloud sparsifying methods, which determines redundant parameters in the pretrained model that lead to minimal distortion in both locality and confidence (detection distortion); and (2) a universal plug-and-play pruning framework that works with arbitrary 3D detection model. This framework aims to minimize detection distortion of network output to maximally maintain detection precision, by identifying layer-wise sparsity based on second-order Taylor approximation of the distortion. Albeit utilizing second-order information, we introduced a lightweight scheme to efficiently acquire Hessian information, and subsequently perform dynamic programming to solve the layer-wise sparsity. Extensive experiments on KITTI, Nuscenes and ONCE datasets demonstrate that our approach is able to maintain and even boost the detection precision on pruned model under noticeable computation reduction (FLOPs). Noticeably, we achieve over 3.89x, 3.72x FLOPs reduction on CenterPoint and PVRCNN model, respectively, without mAP decrease, significantly improving the state-of-the-art.

Jing Wang, Fengzhuo Zhang, Xiaoli Li et al.

A variety of Auto-Regressive Video Diffusion Models (ARVDM) have achieved remarkable successes in generating realistic long-form videos. However, theoretical analyses of these models remain scant. In this work, we develop theoretical underpinnings for these models and use our insights to improve the performance of existing models. We first develop Meta-ARVDM, a unified framework of ARVDMs that subsumes most existing methods. Using Meta-ARVDM, we analyze the KL-divergence between the videos generated by Meta-ARVDM and the true videos. Our analysis uncovers two important phenomena inherent to ARVDM -- error accumulation and memory bottleneck. By deriving an information-theoretic impossibility result, we show that the memory bottleneck phenomenon cannot be avoided. To mitigate the memory bottleneck, we design various network structures to explicitly use more past frames. We also achieve a significantly improved trade-off between the mitigation of the memory bottleneck and the inference efficiency by compressing the frames. Experimental results on DMLab and Minecraft validate the efficacy of our methods. Our experiments also demonstrate a Pareto-frontier between the error accumulation and memory bottleneck across different methods.

Boxiang Ma, Ru Li, Yuanlong Wang et al.

Driven by vast and diverse textual data, large language models (LLMs) have demonstrated impressive performance across numerous natural language processing (NLP) tasks. Yet, a critical question persists: does their generalization arise from mere memorization of training data or from deep semantic understanding? To investigate this, we propose a bi-perspective evaluation framework to assess LLMs' scenario cognition - the ability to link semantic scenario elements with their arguments in context. Specifically, we introduce a novel scenario-based dataset comprising diverse textual descriptions of fictional facts, annotated with scenario elements. LLMs are evaluated through their capacity to answer scenario-related questions (model output perspective) and via probing their internal representations for encoded scenario elements-argument associations (internal representation perspective). Our experiments reveal that current LLMs predominantly rely on superficial memorization, failing to achieve robust semantic scenario cognition, even in simple cases. These findings expose critical limitations in LLMs' semantic understanding and offer cognitive insights for advancing their capabilities.

Yan Qin, Chau Yuen, Yimin Shao et al.

Capsule network (CapsNet) acts as a promising alternative to the typical convolutional neural network, which is the dominant network to develop the remaining useful life (RUL) estimation models for mechanical equipment. Although CapsNet comes with an impressive ability to represent the entities' hierarchical relationships through a high-dimensional vector embedding, it fails to capture the long-term temporal correlation of run-to-failure time series measured from degraded mechanical equipment. On the other hand, the slow-varying dynamics, which reveals the low-frequency information hidden in mechanical dynamical behaviour, is overlooked in the existing RUL estimation models, limiting the utmost ability of advanced networks. To address the aforementioned concerns, we propose a Slow-varying Dynamics assisted Temporal CapsNet (SD-TemCapsNet) to simultaneously learn the slow-varying dynamics and temporal dynamics from measurements for accurate RUL estimation. First, in light of the sensitivity of fault evolution, slow-varying features are decomposed from normal raw data to convey the low-frequency components corresponding to the system dynamics. Next, the long short-term memory (LSTM) mechanism is introduced into CapsNet to capture the temporal correlation of time series. To this end, experiments conducted on an aircraft engine and a milling machine verify that the proposed SD-TemCapsNet outperforms the mainstream methods. In comparison with CapsNet, the estimation accuracy of the aircraft engine with four different scenarios has been improved by 10.17%, 24.97%, 3.25%, and 13.03% concerning the index root mean squared error, respectively. Similarly, the estimation accuracy of the milling machine has been improved by 23.57% compared to LSTM and 19.54% compared to CapsNet.

Yuecong Xu, Jianfei Yang, Haozhi Cao et al.

While action recognition (AR) has gained large improvements with the introduction of large-scale video datasets and the development of deep neural networks, AR models robust to challenging environments in real-world scenarios are still under-explored. We focus on the task of action recognition in dark environments, which can be applied to fields such as surveillance and autonomous driving at night. Intuitively, current deep networks along with visual enhancement techniques should be able to handle AR in dark environments, however, it is observed that this is not always the case in practice. To dive deeper into exploring solutions for AR in dark environments, we launched the UG2+ Challenge Track 2 (UG2-2) in IEEE CVPR 2021, with a goal of evaluating and advancing the robustness of AR models in dark environments. The challenge builds and expands on top of a novel ARID dataset, the first dataset for the task of dark video AR, and guides models to tackle such a task in both fully and semi-supervised manners. Baseline results utilizing current AR models and enhancement methods are reported, justifying the challenging nature of this task with substantial room for improvements. Thanks to the active participation from the research community, notable advances have been made in participants' solutions, while analysis of these solutions helped better identify possible directions to tackle the challenge of AR in dark environments.

Peisheng Qian, Ziyuan Zhao, Cong Chen et al.

Diabetic retinopathy (DR) is one of the most common eye conditions among diabetic patients. However, vision loss occurs primarily in the late stages of DR, and the symptoms of visual impairment, ranging from mild to severe, can vary greatly, adding to the burden of diagnosis and treatment in clinical practice. Deep learning methods based on retinal images have achieved remarkable success in automatic DR grading, but most of them neglect that the presence of diabetes usually affects both eyes, and ophthalmologists usually compare both eyes concurrently for DR diagnosis, leaving correlations between left and right eyes unexploited. In this study, simulating the diagnostic process, we propose a two-stream binocular network to capture the subtle correlations between left and right eyes, in which, paired images of eyes are fed into two identical subnetworks separately during training. We design a contrastive grading loss to learn binocular correlation for five-class DR detection, which maximizes inter-class dissimilarity while minimizing the intra-class difference. Experimental results on the EyePACS dataset show the superiority of the proposed binocular model, outperforming monocular methods by a large margin.

Xiaoli Li

Dropout is attracting intensive research interest in deep learning as an efficient approach to prevent overfitting. Recently incorporating structural information when deciding which units to drop out produced promising results comparing to methods that ignore the structural information. However, a major issue of the existing work is that it failed to differentiate among instances when constructing the dropout architecture. This can be a significant deficiency for many applications. To solve this issue, we propose Constructivism learning for instance-dependent Dropout Architecture (CODA), which is inspired from a philosophical theory, constructivism learning. Specially, based on the theory we have designed a better drop out technique, Uniform Process Mixture Models, using a Bayesian nonparametric method Uniform process. We have evaluated our proposed method on 5 real-world datasets and compared the performance with other state-of-the-art dropout techniques. The experimental results demonstrated the effectiveness of CODA.

Mohamed Ragab, Zhenghua Chen, Min Wu et al.

Accurate estimation of remaining useful life (RUL) of industrial equipment can enable advanced maintenance schedules, increase equipment availability and reduce operational costs. However, existing deep learning methods for RUL prediction are not completely successful due to the following two reasons. First, relying on a single objective function to estimate the RUL will limit the learned representations and thus affect the prediction accuracy. Second, while longer sequences are more informative for modelling the sensor dynamics of equipment, existing methods are less effective to deal with very long sequences, as they mainly focus on the latest information. To address these two problems, we develop a novel attention-based sequence to sequence with auxiliary task (ATS2S) model. In particular, our model jointly optimizes both reconstruction loss to empower our model with predictive capabilities (by predicting next input sequence given current input sequence) and RUL prediction loss to minimize the difference between the predicted RUL and actual RUL. Furthermore, to better handle longer sequence, we employ the attention mechanism to focus on all the important input information during training process. Finally, we propose a new dual-latent feature representation to integrate the encoder features and decoder hidden states, to capture rich semantic information in data. We conduct extensive experiments on four real datasets to evaluate the efficacy of the proposed method. Experimental results show that our proposed method can achieve superior performance over 13 state-of-the-art methods consistently.

Taiping Zeng, Xiaoli Li, Bailu Si

We propose a neurobiologically inspired visual simultaneous localization and mapping (SLAM) system based on direction sparse method to real-time build cognitive maps of large-scale environments from a moving stereo camera. The core SLAM system mainly comprises a Bayesian attractor network, which utilizes neural responses of head direction (HD) cells in the hippocampus and grid cells in the medial entorhinal cortex (MEC) to represent the head direction and the position of the robot in the environment, respectively. Direct sparse method is employed to accurately and robustly estimate velocity information from a stereo camera. Input rotational and translational velocities are integrated by the HD cell and grid cell networks, respectively. We demonstrated our neurobiologically inspired stereo visual SLAM system on the KITTI odometry benchmark datasets. Our proposed SLAM system is robust to real-time build a coherent semi-metric topological map from a stereo camera. Qualitative evaluation on cognitive maps shows that our proposed neurobiologically inspired stereo visual SLAM system outperforms our previous brain-inspired algorithms and the neurobiologically inspired monocular visual SLAM system both in terms of tracking accuracy and robustness, which is closer to the traditional state-of-the-art one.

Taiping Zeng, XiaoLi Li, Bailu Si

The entorhinal-hippocampal circuit plays a critical role in higher brain functions, especially spatial cognition. Grid cells in the medial entorhinal cortex (MEC) periodically fire with different grid spacing and orientation, which makes a contribution that place cells in the hippocampus can uniquely encode locations in an environment. But how sparse firing granule cells in the dentate gyrus are formed from grid cells in the MEC remains to be determined. Recently, the fruit fly olfactory circuit provides a variant algorithm (called locality-sensitive hashing) to solve this problem. To investigate how the sparse place firing generates in the dentate gyrus can help animals to break the perception ambiguity during environment exploration, we build a biologically relevant, computational model from grid cells to place cells. The weight from grid cells to dentate gyrus granule cells is learned by competitive Hebbian learning. We resorted to the robot system for demonstrating our cognitive mapping model on the KITTI odometry benchmark dataset. The experimental results show that our model is able to stably, robustly build a coherent semi-metric topological map in the large-scale outdoor environment. The experimental results suggest that the entorhinal-hippocampal circuit as a variant locality-sensitive hashing algorithm is capable of generating sparse encoding for easily distinguishing different locations in the environment. Our experiments also provide theoretical supports that this analogous hashing algorithm may be a general principle of computation in different brain regions and species.

Bo Zhang, Wei Li, Jie Hao et al.

Traditional intelligent fault diagnosis of rolling bearings work well only under a common assumption that the labeled training data (source domain) and unlabeled testing data (target domain) are drawn from the same distribution. However, in many real-world applications, this assumption does not hold, especially when the working condition varies. In this paper, a new adversarial adaptive 1-D CNN called A2CNN is proposed to address this problem. A2CNN consists of four parts, namely, a source feature extractor, a target feature extractor, a label classifier and a domain discriminator. The layers between the source and target feature extractor are partially untied during the training stage to take both training efficiency and domain adaptation into consideration. Experiments show that A2CNN has strong fault-discriminative and domain-invariant capacity, and therefore can achieve high accuracy under different working conditions. We also visualize the learned features and the networks to explore the reasons behind the high performance of our proposed model.

Gaurav Manek, Jie Lin, Vijay Chandrasekhar et al.

In this work, we focus on the problem of image instance retrieval with deep descriptors extracted from pruned Convolutional Neural Networks (CNN). The objective is to heavily prune convolutional edges while maintaining retrieval performance. To this end, we introduce both data-independent and data-dependent heuristics to prune convolutional edges, and evaluate their performance across various compression rates with different deep descriptors over several benchmark datasets. Further, we present an end-to-end framework to fine-tune the pruned network, with a triplet loss function specially designed for the retrieval task. We show that the combination of heuristic pruning and fine-tuning offers 5x compression rate without considerable loss in retrieval performance.

Honghai Yu, Pierre Moulin, Hong Wei Ng et al.

Hashing has emerged as a popular technique for large-scale similarity search. Most learning-based hashing methods generate compact yet correlated hash codes. However, this redundancy is storage-inefficient. Hence we propose a lossless variable-length hashing (VLH) method that is both storage- and search-efficient. Storage efficiency is achieved by converting the fixed-length hash code into a variable-length code. Search efficiency is obtained by using a multiple hash table structure. With VLH, we are able to deliberately add redundancy into hash codes to improve retrieval performance with little sacrifice in storage efficiency or search complexity. In particular, we propose a block K-means hashing (B-KMH) method to obtain significantly improved retrieval performance with no increase in storage and marginal increase in computational cost.