A5101675123- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Anomaly Detection Techniques and Applications
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Topic Modeling
- Advanced Vision and Imaging
- Advanced Image and Video Retrieval Techniques
- Human Pose and Action Recognition
- Data Stream Mining Techniques
- Data Mining Algorithms and Applications
- Image and Object Detection Techniques
- Traffic Prediction and Management Techniques
- Software System Performance and Reliability
- Adversarial Robustness in Machine Learning
- Time Series Analysis and Forecasting
- Cloud Computing and Resource Management
- Speech Recognition and Synthesis
- Data Management and Algorithms
- 3D Shape Modeling and Analysis
- Medical Image Segmentation Techniques
- Video Analysis and Summarization
- Morphological variations and asymmetry
- Advanced Sensor and Control Systems
- Virtual Reality Applications and Impacts
Tsinghua University
2019-2025
Southwest Petroleum University
2023
University of Wisconsin–Madison
2021
Huawei Technologies (China)
2021
Alibaba Group (United States)
2019
Zhejiang University
2019
Sichuan Agricultural University
2019
Nanjing University of Posts and Telecommunications
2018
Xi'an High Tech University
2018
China University of Geosciences (Beijing)
2016
Detecting out-of-distribution (OOD) inputs is a central challenge for safely deploying machine learning models in the real world. Existing solutions are mainly driven by small datasets, with low resolution and very few class labels (e.g., CIFAR). As result, OOD detection large-scale image classification tasks remains largely unexplored. In this paper, we bridge critical gap proposing group-based framework, along novel scoring function termed MOS. Our key idea to decompose large semantic...
The accuracy of deep convolutional neural networks (CNNs) generally improves when fueled with high resolution images. However, this often comes at a computational cost and memory footprint. Inspired by the fact that not all regions in an image are task-relevant, we propose novel framework performs efficient classification processing sequence relatively small inputs, which strategically selected from original reinforcement learning. Such dynamic decision process naturally facilitates adaptive...
Detecting out-of-distribution (OOD) data has become a critical component in ensuring the safe deployment of machine learning models real world. Existing OOD detection approaches primarily rely on output or feature space for deriving scores, while largely overlooking information from gradient space. In this paper, we present GradNorm, simple and effective approach detecting inputs by utilizing extracted GradNorm directly employs vector norm gradients, backpropagated KL divergence between...
Vision Transformers (ViT) have achieved remarkable success in large-scale image recognition. They split every 2D into a fixed number of patches, each which is treated as token. Generally, representing an with more tokens would lead to higher prediction accuracy, while it also results drastically increased computational cost. To achieve decent trade-off between accuracy and speed, the empirically set 16x16 or 14x14. In this paper, we argue that has its own characteristics, ideally token...
Stereo matching is a key technique for metric depth estimation in computer vision and robotics. Real-world challenges like occlusion non-texture hinder accurate disparity from binocular cues. Recently, monocular relative has shown remarkable generalization using foundation models. Thus, to facilitate robust stereo with cues, we incorporate model into the recurrent stereo-matching framework, building new framework model-based stereo-matching, DEFOM-Stereo. In feature extraction stage,...
Imaging sonar is a crucial tool for underwater visual perception. Compared to 2D images, 3D images offer superior spatial positioning capabilities, although the data acquisition cost higher and lacks open source references annotation, target detection, semantic segmentation. This paper utilizes imaging collect from three types of targets with 1534 effective frames, including tire, mannequin, table, in Liquan Lake, Shanxi Province, China. Based on these data, this study focuses innovative...
The issue of high system stability is one the major obstacles for real-time computing over fluctuating big data streams. A stable scheduling more important than an efficient stream applications, especially when a to be rescheduled dynamically at runtime. In this paper, online strategy with makespan guarantee SOMG discussed, which includes following features: 1) profiling mathematical relationships between stability, response time, and resource utilization, indicating conditions meet...
Code completion is widely used by software developers to provide coding suggestions given a partially written code snippet. Apart from the traditional methods, which only support single token at minimal positions, recent studies show ability longer more flexible positions. However, such frequently triggered and results reduce overall precision as they generate invalid results. Moreover, different are mostly incompatible with each other. Thus, it vital develop an ensemble framework that can...
Intersections are the key to improve traffic efficiency. For intersections with complex conditions, if we want efficiency effectively, should make signals adjust adaptively according different status. Obviously, traditional fixed timing strategy is hard achieve this. In addition, cooperative control of multiple will maximize their overall interests and reduce contradictions between intersections. Therefore, in this paper, propose an adaptive signal method for based on deep reinforcement...
Stream computing systems process high-rate incoming data sources in dynamic environments and generate real-time results. A critical problem stream processing is the difficulty devising an optimal strategy for configuring topology structure. That due to inability predict or adapt dynamics of flow intensity platform's workload. Therefore, best should rely on statistics performance workload achieve self-adaptation. In this paper, we first analyse impact application structures Storm resource...
Virtual worlds have the potential to enable and enhance online learning outcomes. Because in three-dimensional (3D) designed spaces depends on learners’ spatial processing abilities, we need understand how these abilities may affect Building hunter-gatherer theory of gender difference examined interacts with type (directed vs. incidental) virtual world (VR) simulations objects. Specifically, theorized that men’s women’s would lead differential outcomes based instructor designed. Using a...
In recent years, with the significant evolution of multi-modal large models, many recommender researchers realized potential information for user interest modeling. industry, a wide-used modeling architecture is cascading paradigm: (1) first pre-training model to provide omnipotent representations downstream services; (2) The recommendation takes representation as additional input fit real user-item behaviours. Although such paradigm achieves remarkable improvements, however, there still...