Yu Hen Hu

Cheng-En Wu, Yu Tian, Haichao Yu et al.

Vision-language models such as CLIP learn a generic text-image embedding from large-scale training data. A vision-language model can be adapted to a new classification task through few-shot prompt tuning. We find that such a prompt tuning process is highly robust to label noises. This intrigues us to study the key reasons contributing to the robustness of the prompt tuning paradigm. We conducted extensive experiments to explore this property and find the key factors are: 1) the fixed classname tokens provide a strong regularization to the optimization of the model, reducing gradients induced by the noisy samples; 2) the powerful pre-trained image-text embedding that is learned from diverse and generic web data provides strong prior knowledge for image classification. Further, we demonstrate that noisy zero-shot predictions from CLIP can be used to tune its own prompt, significantly enhancing prediction accuracy in the unsupervised setting. The code is available at https://github.com/CEWu/PTNL.

Zhenyu Yang, Yuhang Hu, Zemin Du et al.

Despite the significant advancements of Large Vision-Language Models (LVLMs) on established benchmarks, there remains a notable gap in suitable evaluation regarding their applicability in the emerging domain of long-context streaming video understanding. Current benchmarks for video understanding typically emphasize isolated single-instance text inputs and fail to evaluate the capacity to sustain temporal reasoning throughout the entire duration of video streams. To address these limitations, we introduce SVBench, a pioneering benchmark with temporal multi-turn question-answering chains specifically designed to thoroughly assess the capabilities of streaming video understanding of current LVLMs. We design a semi-automated annotation pipeline to obtain 49,979 Question-Answer (QA) pairs of 1,353 streaming videos, which includes generating QA chains that represent a series of consecutive multi-turn dialogues over video segments and constructing temporal linkages between successive QA chains. Our experimental results, obtained from 14 models in dialogue and streaming evaluations, reveal that while the closed-source GPT-4o outperforms others, most open-source LVLMs struggle with long-context streaming video understanding. We also construct a StreamingChat model, which significantly outperforms open-source LVLMs on our SVBench and achieves comparable performance on diverse vision-language benchmarks. We expect SVBench to advance the research of streaming video understanding by providing a comprehensive and in-depth analysis of current LVLMs. Our benchmark and model can be accessed at https://github.com/sotayang/SVBench.

Ji Qiu, Hongmei Shi, Yu Hen Hu et al.

Frequent false alarms impede the promotion of unsupervised anomaly detection algorithms in industrial applications. Potential characteristics of false alarms depending on the trained detector are revealed by investigating density probability distributions of prediction scores in the out-of-distribution anomaly detection tasks. An SVM-based classifier is exploited as a post-processing module to identify false alarms from the anomaly map at the object level. Besides, a sample synthesis strategy is devised to incorporate fuzzy prior knowledge on the specific application in the anomaly-free training dataset. Experimental results illustrate that the proposed method comprehensively improves the performances of two segmentation models at both image and pixel levels on two industrial applications.

Cheng-En Wu, Farley Lai, Yu Hen Hu et al.

Self-supervised video representation learning has been shown to effectively improve downstream tasks such as video retrieval and action recognition. In this paper, we present the Cascade Positive Retrieval (CPR) that successively mines positive examples w.r.t. the query for contrastive learning in a cascade of stages. Specifically, CPR exploits multiple views of a query example in different modalities, where an alternative view may help find another positive example dissimilar in the query view. We explore the effects of possible CPR configurations in ablations including the number of mining stages, the top similar example selection ratio in each stage, and progressive training with an incremental number of the final Top-k selection. The overall mining quality is measured to reflect the recall across training set classes. CPR reaches a median class mining recall of 83.3%, outperforming previous work by 5.5%. Implementation-wise, CPR is complementary to pretext tasks and can be easily applied to previous work. In the evaluation of pretraining on UCF101, CPR consistently improves existing work and even achieves state-of-the-art R@1 of 56.7% and 24.4% in video retrieval as well as 83.8% and 54.8% in action recognition on UCF101 and HMDB51. The code is available at https://github.com/necla-ml/CPR.

Cheng-En Wu, Jinhong Lin, Yu Hen Hu et al.

Contrastive image-text pre-trained models such as CLIP have shown remarkable adaptability to downstream tasks. However, they face challenges due to the high computational requirements of the Vision Transformer (ViT) backbone. Current strategies to boost ViT efficiency focus on pruning patch tokens but fall short in addressing the multimodal nature of CLIP and identifying the optimal subset of tokens for maximum performance. To address this, we propose greedy search methods to establish a "Golden Ranking" and introduce a lightweight predictor specifically trained to approximate this Ranking. To compensate for any performance degradation resulting from token pruning, we incorporate learnable visual tokens that aid in restoring and potentially enhancing the model's performance. Our work presents a comprehensive and systematic investigation of pruning tokens within the ViT backbone of CLIP models. Through our framework, we successfully reduced 40% of patch tokens in CLIP's ViT while only suffering a minimal average accuracy loss of 0.3 across seven datasets. Our study lays the groundwork for building more computationally efficient multimodal models without sacrificing their performance, addressing a key challenge in the application of advanced vision-language models.

Jiangang Chen, Yung-Hong Sun, Kristen Pickett et al.

We developed a shoe-mounted gait monitoring system capable of tracking up to 17 gait parameters, including gait length, step time, stride velocity, and others. The system employs a stereo camera mounted on one shoe to track a marker placed on the opposite shoe, enabling the estimation of spatial gait parameters. Additionally, a Force Sensitive Resistor (FSR) affixed to the heel of the shoe, combined with a custom-designed algorithm, is utilized to measure temporal gait parameters. Through testing on multiple participants and comparison with the gait mat, the proposed gait monitoring system exhibited notable performance, with the accuracy of all measured gait parameters exceeding 93.61%. The system also demonstrated a low drift of 4.89% during long-distance walking. A gait identification task conducted on participants using a trained Transformer model achieved 95.7% accuracy on the dataset collected by the proposed system, demonstrating that our hardware has the potential to collect long-sequence gait data suitable for integration with current Large Language Models (LLMs). The system is cost-effective, user-friendly, and well-suited for real-life measurements.

Gefei Shen, Yung-Hong Sun, Yu Hen Hu et al.

Two sampling strategies are investigated to enhance efficiency in training a deep learning object detection model. These sampling strategies are employed under the assumption of Lipschitz continuity of deep learning models. The first strategy is uniform sampling which seeks to obtain samples evenly yet randomly through the state space of the object dynamics. The second strategy of frame difference sampling is developed to explore the temporal redundancy among successive frames in a video. Experiment result indicates that these proposed sampling strategies provide a dataset that yields good training performance while requiring relatively few manually labelled samples.

Kyle Boone, Ben Wurster, Seth Thao et al.

This project is centered around building a neural network that is able to recognize ASL letters in images, particularly within the scope of a live video feed. Initial testing results came up short of expectations when both the convolutional network and VGG16 transfer learning approaches failed to generalize in settings of different backgrounds. The use of a pre-trained hand joint detection model was then adopted with the produced joint locations being fed into a fully-connected neural network. The results of this approach exceeded those of prior methods and generalized well to a live video feed application.

Zhengyang Lou, Huan Xu, Fangzhou Mu et al.

Deep models have demonstrated recent success in single-image dehazing. Most prior methods consider fully supervised training and learn from paired clean and hazy images, where a hazy image is synthesized based on a clean image and its estimated depth map. This paradigm, however, can produce low-quality hazy images due to inaccurate depth estimation, resulting in poor generalization of the trained models. In this paper, we explore an alternative approach for generating paired clean-hazy images by leveraging computer graphics. Using a modern game engine, our approach renders crisp clean images and their precise depth maps, based on which high-quality hazy images can be synthesized for training dehazing models. To this end, we present SimHaze: a new synthetic haze dataset. More importantly, we show that training with SimHaze alone allows the latest dehazing models to achieve significantly better performance in comparison to previous dehazing datasets. Our dataset and code will be made publicly available.

Wentao Chen, Hailong Qiu, Jian Zhuang et al.

Deep neural networks (DNNs) have demonstrated their great potential in recent years, exceeding the per-formance of human experts in a wide range of applications. Due to their large sizes, however, compressiontechniques such as weight quantization and pruning are usually applied before they can be accommodated onthe edge. It is generally believed that quantization leads to performance degradation, and plenty of existingworks have explored quantization strategies aiming at minimum accuracy loss. In this paper, we argue thatquantization, which essentially imposes regularization on weight representations, can sometimes help toimprove accuracy. We conduct comprehensive experiments on three widely used applications: fully con-nected network (FCN) for biomedical image segmentation, convolutional neural network (CNN) for imageclassification on ImageNet, and recurrent neural network (RNN) for automatic speech recognition, and experi-mental results show that quantization can improve the accuracy by 1%, 1.95%, 4.23% on the three applicationsrespectively with 3.5x-6.4x memory reduction.

Run-Ze Wang, Zhen-Hua Ling, Jing-Bo Zhou et al.

The task of multi-turn text-to-SQL semantic parsing aims to translate natural language utterances in an interaction into SQL queries in order to answer them using a database which normally contains multiple table schemas. Previous studies on this task usually utilized contextual information to enrich utterance representations and to further influence the decoding process. While they ignored to describe and track the interaction states which are determined by history SQL queries and are related with the intent of current utterance. In this paper, two kinds of interaction states are defined based on schema items and SQL keywords separately. A relational graph neural network and a non-linear layer are designed to update the representations of these two states respectively. The dynamic schema-state and SQL-state representations are then utilized to decode the SQL query corresponding to current utterance. Experimental results on the challenging CoSQL dataset demonstrate the effectiveness of our proposed method, which achieves better performance than other published methods on the task leaderboard.

Qiankun Tang, Shice Liu, Jie Li et al.

We introduce a novel single-shot object detector to ease the imbalance of foreground-background class by suppressing the easy negatives while increasing the positives. To achieve this, we propose an Anchor Promotion Module (APM) which predicts the probability of each anchor as positive and adjusts their initial locations and shapes to promote both the quality and quantity of positive anchors. In addition, we design an efficient Feature Alignment Module (FAM) to extract aligned features for fitting the promoted anchors with the help of both the location and shape transformation information from the APM. We assemble the two proposed modules to the backbone of VGG-16 and ResNet-101 network with an encoder-decoder architecture. Extensive experiments on MS COCO well demonstrate our model performs competitively with alternative methods (40.0\% mAP on \textit{test-dev} set) and runs faster (28.6 \textit{fps}).

Boyu Zhang, Azadeh Davoodi, Yu Hen Hu

The deployment of Convolutional Neural Networks (CNNs) on resource constrained platforms such as mobile devices and embedded systems has been greatly hindered by their high implementation cost, and thus motivated a lot research interest in compressing and accelerating trained CNN models. Among various techniques proposed in literature, structured pruning, especially channel pruning, has gain a lot focus due to 1) its superior performance in memory, computation, and energy reduction; and 2) it is friendly to existing hardware and software libraries. In this paper, we investigate the intermediate results of convolutional layers and present a novel pivoted QR factorization based channel pruning technique that can prune any specified number of input channels of any layer. We also explore more pruning opportunities in ResNet-like architectures by applying two tweaks to our technique. Experiment results on VGG-16 and ResNet-50 models with ImageNet ILSVRC 2012 dataset are very impressive with 4.29X and 2.84X computation reduction while only sacrificing about 1.40\% top-5 accuracy. Compared to many prior works, the pruned models produced by our technique require up to 47.7\% less computation while still achieve higher accuracies.

Boyu Zhang, Azadeh Davoodi, Yu-Hen Hu

The ability to customize a trained Deep Neural Network (DNN) locally using user-specific data may greatly enhance user experiences, reduce development costs, and protect user's privacy. In this work, we propose to incorporate a novel Mixture of Experts (MOE) approach to accomplish this goal. This architecture comprises of a Global Expert (GE), a Local Expert (LE) and a Gating Network (GN). The GE is a trained DNN developed on a large training dataset representative of many potential users. After deployment on an embedded edge device, GE will be subject to customized, user-specific data (e.g., accent in speech) and its performance may suffer. This problem may be alleviated by training a local DNN (the local expert, LE) on a small size customized training data to correct the errors made by GE. A gating network then will be trained to determine whether an incoming data should be handled by GE or LE. Since the customized dataset is in general very small, the cost of training LE and GN would be much lower than that of re-training of GE. The training of LE and GN thus can be performed at local device, properly protecting the privacy of customized training data. In this work, we developed a prototype MOE architecture for handwritten alphanumeric character recognition task. We use EMNIST as the generic dataset, LeNet5 as GE, and handwritings of 10 users as the customized dataset. We show that with the LE and GN, the classification accuracy is significantly enhanced over the customized dataset with almost no degradation of accuracy over the generic dataset. In terms of energy and network size, the overhead of LE and GN is around 2.5% compared to those of GE.

Xiaowei Xu, Qing Lu, Yu Hu et al.

With pervasive applications of medical imaging in health-care, biomedical image segmentation plays a central role in quantitative analysis, clinical diagno- sis, and medical intervention. Since manual anno- tation su ers limited reproducibility, arduous e orts, and excessive time, automatic segmentation is desired to process increasingly larger scale histopathological data. Recently, deep neural networks (DNNs), par- ticularly fully convolutional networks (FCNs), have been widely applied to biomedical image segmenta- tion, attaining much improved performance. At the same time, quantization of DNNs has become an ac- tive research topic, which aims to represent weights with less memory (precision) to considerably reduce memory and computation requirements of DNNs while maintaining acceptable accuracy. In this paper, we apply quantization techniques to FCNs for accurate biomedical image segmentation. Unlike existing litera- ture on quantization which primarily targets memory and computation complexity reduction, we apply quan- tization as a method to reduce over tting in FCNs for better accuracy. Speci cally, we focus on a state-of- the-art segmentation framework, suggestive annotation [22], which judiciously extracts representative annota- tion samples from the original training dataset, obtain- ing an e ective small-sized balanced training dataset. We develop two new quantization processes for this framework: (1) suggestive annotation with quantiza- tion for highly representative training samples, and (2) network training with quantization for high accuracy. Extensive experiments on the MICCAI Gland dataset show that both quantization processes can improve the segmentation performance, and our proposed method exceeds the current state-of-the-art performance by up to 1%. In addition, our method has a reduction of up to 6.4x on memory usage.

Quan Liu, Hui Jiang, Zhen-Hua Ling et al.

In this paper, we propose commonsense knowledge enhanced embeddings (KEE) for solving the Pronoun Disambiguation Problems (PDP). The PDP task we investigate in this paper is a complex coreference resolution task which requires the utilization of commonsense knowledge. This task is a standard first round test set in the 2016 Winograd Schema Challenge. In this task, traditional linguistic features that are useful for coreference resolution, e.g. context and gender information, are no longer effective anymore. Therefore, the KEE models are proposed to provide a general framework to make use of commonsense knowledge for solving the PDP problems. Since the PDP task doesn't have training data, the KEE models would be used during the unsupervised feature extraction process. To evaluate the effectiveness of the KEE models, we propose to incorporate various commonsense knowledge bases, including ConceptNet, WordNet, and CauseCom, into the KEE training process. We achieved the best performance by applying the proposed methods to the 2016 Winograd Schema Challenge. In addition, experiments conducted on the standard PDP task indicate that, the proposed KEE models could solve the PDP problems by achieving 66.7% accuracy, which is a new state-of-the-art performance.