Gang Chen

Jie Song, Zhengqi Xu, Sai Wu et al.

The last decade has witnessed the success of deep learning and the surge of publicly released trained models, which necessitates the quantification of the model functional distance for various purposes. However, quantifying the model functional distance is always challenging due to the opacity in inner workings and the heterogeneity in architectures or tasks. Inspired by the concept of "field" in physics, in this work we introduce Model Gradient Field (abbr. ModelGiF) to extract homogeneous representations from the heterogeneous pre-trained models. Our main assumption underlying ModelGiF is that each pre-trained deep model uniquely determines a ModelGiF over the input space. The distance between models can thus be measured by the similarity between their ModelGiFs. We validate the effectiveness of the proposed ModelGiF with a suite of testbeds, including task relatedness estimation, intellectual property protection, and model unlearning verification. Experimental results demonstrate the versatility of the proposed ModelGiF on these tasks, with significantly superiority performance to state-of-the-art competitors. Codes are available at https://github.com/zju-vipa/modelgif.

Hanyu Zhou, Yi Chang, Gang Chen et al.

Optical flow estimation has made great progress, but usually suffers from degradation under adverse weather. Although semi/full-supervised methods have made good attempts, the domain shift between the synthetic and real adverse weather images would deteriorate their performance. To alleviate this issue, our start point is to unsupervisedly transfer the knowledge from source clean domain to target degraded domain. Our key insight is that adverse weather does not change the intrinsic optical flow of the scene, but causes a significant difference for the warp error between clean and degraded images. In this work, we propose the first unsupervised framework for adverse weather optical flow via hierarchical motion-boundary adaptation. Specifically, we first employ image translation to construct the transformation relationship between clean and degraded domains. In motion adaptation, we utilize the flow consistency knowledge to align the cross-domain optical flows into a motion-invariance common space, where the optical flow from clean weather is used as the guidance-knowledge to obtain a preliminary optical flow for adverse weather. Furthermore, we leverage the warp error inconsistency which measures the motion misalignment of the boundary between the clean and degraded domains, and propose a joint intra- and inter-scene boundary contrastive adaptation to refine the motion boundary. The hierarchical motion and boundary adaptation jointly promotes optical flow in a unified framework. Extensive quantitative and qualitative experiments have been performed to verify the superiority of the proposed method.

Haitao Meng, Changcai Li, Gang Chen et al.

The proposal of Pseudo-Lidar representation has significantly narrowed the gap between visual-based and active Lidar-based 3D object detection. However, current researches exclusively focus on pushing the accuracy improvement of Pseudo-Lidar by taking the advantage of complex and time-consuming neural networks. Seldom explore the profound characteristics of Pseudo-Lidar representation to obtain the promoting opportunities. In this paper, we dive deep into the pseudo Lidar representation and argue that the performance of 3D object detection is not fully dependent on the high precision stereo depth estimation. We demonstrate that even for the unreliable depth estimation, with proper data processing and refining, it can achieve comparable 3D object detection accuracy. With this finding, we further show the possibility that utilizing fast but inaccurate stereo matching algorithms in the Pseudo-Lidar system to achieve low latency responsiveness. In the experiments, we develop a system with a less powerful stereo matching predictor and adopt the proposed refinement schemes to improve the accuracy. The evaluation on the KITTI benchmark shows that the presented system achieves competitive accuracy to the state-of-the-art approaches with only 23 ms computing, showing it is a suitable candidate for deploying to real car-hold applications.

Hanyu Zhou, Yi Chang, Zhiwei Shi et al.

Optical flow has made great progress in clean scenes, while suffers degradation under adverse weather due to the violation of the brightness constancy and gradient continuity assumptions of optical flow. Typically, existing methods mainly adopt domain adaptation to transfer motion knowledge from clean to degraded domain through one-stage adaptation. However, this direct adaptation is ineffective, since there exists a large gap due to adverse weather and scene style between clean and real degraded domains. Moreover, even within the degraded domain itself, static weather (e.g., fog) and dynamic weather (e.g., rain) have different impacts on optical flow. To address above issues, we explore synthetic degraded domain as an intermediate bridge between clean and real degraded domains, and propose a cumulative homogeneous-heterogeneous adaptation framework for real adverse weather optical flow. Specifically, for clean-degraded transfer, our key insight is that static weather possesses the depth-association homogeneous feature which does not change the intrinsic motion of the scene, while dynamic weather additionally introduces the heterogeneous feature which results in a significant boundary discrepancy in warp errors between clean and degraded domains. For synthetic-real transfer, we figure out that cost volume correlation shares a similar statistical histogram between synthetic and real degraded domains, benefiting to holistically aligning the homogeneous correlation distribution for synthetic-real knowledge distillation. Under this unified framework, the proposed method can progressively and explicitly transfer knowledge from clean scenes to real adverse weather. In addition, we further collect a real adverse weather dataset with manually annotated optical flow labels and perform extensive experiments to verify the superiority of the proposed method.

Gang Chen, Yawei Li, Sargur N. Srihari

Visual restoration and recognition are traditionally addressed in pipeline fashion, i.e. denoising followed by classification. Instead, observing correlations between the two tasks, for example clearer image will lead to better categorization and vice visa, we propose a joint framework for visual restoration and recognition for handwritten images, inspired by advances in deep autoencoder and multi-modality learning. Our model is a 3-pathway deep architecture with a hidden-layer representation which is shared by multi-inputs and outputs, and each branch can be composed of a multi-layer deep model. Thus, visual restoration and classification can be unified using shared representation via non-linear mapping, and model parameters can be learnt via backpropagation. Using MNIST and USPS data corrupted with structured noise, the proposed framework performs at least 20\% better in classification than separate pipelines, as well as clearer recovered images. The noise model and the reproducible source code is available at {\url{https://github.com/ganggit/jointmodel}}.

Gang Chen, Yawei Li, Sargur N. Srihari

Recognition of handwritten words continues to be an important problem in document analysis and recognition. Existing approaches extract hand-engineered features from word images--which can perform poorly with new data sets. Recently, deep learning has attracted great attention because of the ability to learn features from raw data. Moreover they have yielded state-of-the-art results in classification tasks including character recognition and scene recognition. On the other hand, word recognition is a sequential problem where we need to model the correlation between characters. In this paper, we propose using deep Conditional Random Fields (deep CRFs) for word recognition. Basically, we combine CRFs with deep learning, in which deep features are learned and sequences are labeled in a unified framework. We pre-train the deep structure with stacked restricted Boltzmann machines (RBMs) for feature learning and optimize the entire network with an online learning algorithm. The proposed model was evaluated on two datasets, and seen to perform significantly better than competitive baseline models. The source code is available at https://github.com/ganggit/deepCRFs.

Babar Hussain, Qiang Liu, Gang Chen et al.

This paper presents an AI-assisted auto-labeling system for display panel defect detection that leverages in-context learning capabilities. We adopt and enhance the SegGPT architecture with several domain-specific training techniques and introduce a scribble-based annotation mechanism to streamline the labeling process. Our two-stage training approach, validated on industrial display panel datasets, demonstrates significant improvements over the baseline model, achieving an average IoU increase of 0.22 and a 14% improvement in recall across multiple product types, while maintaining approximately 60% auto-labeling coverage. Experimental results show that models trained on our auto-labeled data match the performance of those trained on human-labeled data, offering a practical solution for reducing manual annotation efforts in industrial inspection systems.

Jun Zhang, Jue Wang, Huan Li et al.

We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting. Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM. Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99$\times$.

Gang Chen

The idea of using the recurrent neural network for visual attention has gained popularity in computer vision community. Although the recurrent attention model (RAM) leverages the glimpses with more large patch size to increasing its scope, it may result in high variance and instability. For example, we need the Gaussian policy with high variance to explore object of interests in a large image, which may cause randomized search and unstable learning. In this paper, we propose to unify the top-down and bottom-up attention together for recurrent visual attention. Our model exploits the image pyramids and Q-learning to select regions of interests in the top-down attention mechanism, which in turn to guide the policy search in the bottom-up approach. In addition, we add another two constraints over the bottom-up recurrent neural networks for better exploration. We train our model in an end-to-end reinforcement learning framework, and evaluate our method on visual classification tasks. The experimental results outperform convolutional neural networks (CNNs) baseline and the bottom-up recurrent attention models on visual classification tasks.

Gang Chen, Shengyu He, Haitao Meng et al.

Over the last years, a great success of deep neural networks (DNNs) has been witnessed in computer vision and other fields. However, performance and power constraints make it still challenging to deploy DNNs on mobile devices due to their high computational complexity. Binary neural networks (BNNs) have been demonstrated as a promising solution to achieve this goal by using bit-wise operations to replace most arithmetic operations. Currently, existing GPU-accelerated implementations of BNNs are only tailored for desktop platforms. Due to architecture differences, mere porting of such implementations to mobile devices yields suboptimal performance or is impossible in some cases. In this paper, we propose PhoneBit, a GPU-accelerated BNN inference engine for Android-based mobile devices that fully exploits the computing power of BNNs on mobile GPUs. PhoneBit provides a set of operator-level optimizations including locality-friendly data layout, bit packing with vectorization and layers integration for efficient binary convolution. We also provide a detailed implementation and parallelization optimization for PhoneBit to optimally utilize the memory bandwidth and computing power of mobile GPUs. We evaluate PhoneBit with AlexNet, YOLOv2 Tiny and VGG16 with their binary version. Our experiment results show that PhoneBit can achieve significant speedup and energy efficiency compared with state-of-the-art frameworks for mobile devices.

Gang Chen, Hongzhe Yu, Wei Dong et al.

While training an end-to-end navigation network in the real world is usually of high cost, simulation provides a safe and cheap environment in this training stage. However, training neural network models in simulation brings up the problem of how to effectively transfer the model from simulation to the real world (sim-to-real). In this work, we regard the environment representation as a crucial element in this transfer process and propose a visual information pyramid (VIP) model to systematically investigate a practical environment representation. A novel representation composed of spatial and semantic information synthesis is then established accordingly, where noise model embedding is particularly considered. To explore the effectiveness of this representation, we compared the performance with representations popularly used in the literature in both simulated and real-world scenarios. Results suggest that our environment representation stands out. Furthermore, an analysis on the feature map is implemented to investigate the effectiveness through inner reaction, which could be irradiative for future researches on end-to-end navigation.

Shaofeng Cai, Yao Shu, Gang Chen et al.

Convolutional Neural networks (CNNs) based applications have become ubiquitous, where proper regularization is greatly needed. To prevent large neural network models from overfitting, dropout has been widely used as an efficient regularization technique in practice. However, many recent works show that the standard dropout is ineffective or even detrimental to the training of CNNs. In this paper, we revisit this issue and examine various dropout variants in an attempt to improve existing dropout-based regularization techniques for CNNs. We attribute the failure of standard dropout to the conflict between the stochasticity of dropout and its following Batch Normalization (BN), and propose to reduce the conflict by placing dropout operations right before the convolutional operation instead of BN, or totally address this issue by replacing BN with Group Normalization (GN). We further introduce a structurally more suited dropout variant Drop-Conv2d, which provides more efficient and effective regularization for deep CNNs. These dropout variants can be readily integrated into the building blocks of CNNs and implemented in existing deep learning platforms. Extensive experiments on benchmark datasets including CIFAR, SVHN and ImageNet are conducted to compare the existing building blocks and the proposed ones with dropout training. Results show that our building blocks improve over state-of-the-art CNNs significantly, which is mainly due to the better regularization and implicit model ensemble effect.

Shaofeng Cai, Gang Chen, Beng Chin Ooi et al.

Deep learning models have been used to support analytics beyond simple aggregation, where deeper and wider models have been shown to yield great results. These models consume a huge amount of memory and computational operations. However, most of the large-scale industrial applications are often computational budget constrained. In practice, the peak workload of inference service could be 10x higher than the average cases, with the presence of unpredictable extreme cases. Lots of computational resources could be wasted during off-peak hours and the system may crash when the workload exceeds system capacity. How to support deep learning services with a dynamic workload cost-efficiently remains a challenging problem. In this paper, we address the challenge with a general and novel training scheme called model slicing, which enables deep learning models to provide predictions within the prescribed computational resource budget dynamically. Model slicing could be viewed as an elastic computation solution without requiring more computational resources. Succinctly, each layer in the model is divided into groups of a contiguous block of basic components (i.e. neurons in dense layers and channels in convolutional layers), and then partially ordered relation is introduced to these groups by enforcing that groups participated in each forward pass always starts from the first group to the dynamically-determined rightmost group. Trained by dynamically indexing the rightmost group with a single parameter slice rate, the network is engendered to build up group-wise and residual representation. Then during inference, a sub-model with fewer groups can be readily deployed for efficiency whose computation is roughly quadratic to the width controlled by the slice rate. Extensive experiments show that models trained with model slicing can effectively support on-demand workload with elastic inference cost.

Wei Wang, Gang Chen, Haibo Chen et al.

Recently, deep learning techniques have enjoyed success in various multimedia applications, such as image classification and multi-modal data analysis. Large deep learning models are developed for learning rich representations of complex data. There are two challenges to overcome before deep learning can be widely adopted in multimedia and other applications. One is usability, namely the implementation of different models and training algorithms must be done by non-experts without much effort especially when the model is large and complex. The other is scalability, that is the deep learning system must be able to provision for a huge demand of computing resources for training large models with massive datasets. To address these two challenges, in this paper, we design a distributed deep learning platform called SINGA which has an intuitive programming model based on the common layer abstraction of deep learning models. Good scalability is achieved through flexible distributed training architecture and specific optimization techniques. SINGA runs on GPUs as well as on CPUs, and we show that it outperforms many other state-of-the-art deep learning systems. Our experience with developing and training deep learning models for real-life multimedia applications in SINGA shows that the platform is both usable and scalable.

Gang Chen, Ran Xu, Sargur Srihari

Deep learning has attracted great attention recently and yielded the state of the art performance in dimension reduction and classification problems. However, it cannot effectively handle the structured output prediction, e.g. sequential labeling. In this paper, we propose a deep learning structure, which can learn discriminative features for sequential labeling problems. More specifically, we add the inter-relationship between labels in our deep learning structure, in order to incorporate the context information from the sequential data. Thus, our model is more powerful than linear Conditional Random Fields (CRFs) because the objective function learns latent non-linear features so that target labeling can be better predicted. We pretrain the deep structure with stacked restricted Boltzmann machines (RBMs) for feature learning and optimize our objective function with online learning algorithm, a mixture of perceptron training and stochastic gradient descent. We test our model on different challenge tasks, and show that our model outperforms significantly over the completive baselines.

Gang Chen

Linear Discriminant Analysis (LDA) is a well-known method for dimensionality reduction and classification. Previous studies have also extended the binary-class case into multi-classes. However, many applications, such as object detection and keyframe extraction cannot provide consistent instance-label pairs, while LDA requires labels on instance level for training. Thus it cannot be directly applied for semi-supervised classification problem. In this paper, we overcome this limitation and propose a latent variable Fisher discriminant analysis model. We relax the instance-level labeling into bag-level, is a kind of semi-supervised (video-level labels of event type are required for semantic frame extraction) and incorporates a data-driven prior over the latent variables. Hence, our method combines the latent variable inference and dimension reduction in an unified bayesian framework. We test our method on MUSK and Corel data sets and yield competitive results compared to the baseline approach. We also demonstrate its capacity on the challenging TRECVID MED11 dataset for semantic keyframe extraction and conduct a human-factors ranking-based experimental evaluation, which clearly demonstrates our proposed method consistently extracts more semantically meaningful keyframes than challenging baselines.

Gang Chen, Jason Corso

Multiple instance learning (MIL) has attracted great attention recently in machine learning community. However, most MIL algorithms are very slow and cannot be applied to large datasets. In this paper, we propose a greedy strategy to speed up the multiple instance learning process. Our contribution is two fold. First, we propose a density ratio model, and show that maximizing a density ratio function is the low bound of the DD model under certain conditions. Secondly, we make use of a histogram ratio between positive bags and negative bags to represent the density ratio function and find codebooks separately for positive bags and negative bags by a greedy strategy. For testing, we make use of a nearest neighbor strategy to classify new bags. We test our method on both small benchmark datasets and the large TRECVID MED11 dataset. The experimental results show that our method yields comparable accuracy to the current state of the art, while being up to at least one order of magnitude faster.