A5101719486- Domain Adaptation and Few-Shot Learning
- Advanced Graph Neural Networks
- Augmented Reality Applications
- Topic Modeling
- Interactive and Immersive Displays
- Complex Network Analysis Techniques
- Image Enhancement Techniques
- VLSI and FPGA Design Techniques
- Text Readability and Simplification
- Text and Document Classification Technologies
- Machine Learning and Data Classification
- Functional Brain Connectivity Studies
- Adversarial Robustness in Machine Learning
- Multimodal Machine Learning Applications
- Sparse and Compressive Sensing Techniques
- Advanced Neural Network Applications
- Natural Language Processing Techniques
- Embedded Systems Design Techniques
- Auction Theory and Applications
- Evolutionary Psychology and Human Behavior
- Mobile Crowdsensing and Crowdsourcing
- IoT and Edge/Fog Computing
- Expert finding and Q&A systems
- Parallel Computing and Optimization Techniques
- Virtual Reality Applications and Impacts
University of California, Los Angeles
2015-2024
Chongqing Jiaotong University
2024
Tsinghua University
2022
University of Hong Kong
2014-2016
Hong Kong University of Science and Technology
2014-2016
Background Anorexia nervosa (AN) and body dysmorphic disorder (BDD) are characterized by distorted image frequently co-morbid with each other, although their relationship remains little studied. While there is evidence of abnormalities in visual visuospatial processing both disorders, no study has directly compared the two. We used two complementary modalities – event-related potentials (ERPs) functional magnetic resonance imaging (fMRI) to test for abnormal activity associated early...
We describe a new set of interaction techniques that allow users to interact with physical objects through augmented reality (AR). Previously, operate smart device, touch is generally needed and graphical interface normally involved. These become limitations prevent the user from operating device out reach or multiple devices at once. Ubii (Ubiquitous interaction) an integrated system connects network together, allows using hand gestures. The wears glass which displays in view. Hand gestures...
The recent surge in popularity of crowdsourcing has brought with it a new opportunity for engaging human intelligence the process data analysis. Crowdsourcing provides fundamental mechanism enabling online workers to participate tasks that are either too difficult be solved solely by computer or expensive employ experts perform. In field social science, four elements required form wise crowd - Diversity Opinion, Independence, Decentralization and Aggregation. However, while other three...
Mobile augmented reality (MAR) has exploded in popularity on mobile devices various fields. However, building a MAR application from scratch is complicated and time-consuming. In this paper, we propose CloudRidAR, framework for developers to facilitate the development, deployment, maintenance of applications with little effort. Despite advance as computing platform, their performance still very limited due poor capability devices. order alleviate problem, our CloudRidAR designed cloud at...
We present Ubii (Ubiquitous interface and interaction), an system that aims to expand people's perception interaction from the digital space physical world. The centralized user is broken into pieces woven in domain environment. Augmented paired objects, where presentations are displayed same context. augmented affordance respond as one control provide seamless interaction. By connecting with presents a nearby embodiment afford users sense of awareness interact objects. Integrated on...
The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, applicable Given importance world knowledge LLMs, construct a Knowledge-oriented Assessment benchmark (KoLA), which carefully design three crucial factors: (1) For ability modeling, mimic human cognition form four-level taxonomy knowledge-related...
Unsupervised Source (data) Free domain adaptation (USFDA) aims to transfer knowledge from a well-trained source model related but unlabeled target domain. In such scenario, all conventional methods that require data fail. To combat this challenge, existing USFDAs turn by aligning the feature latent distribution hidden in model. However, information is naturally limited. Thus, alignment scenario not only difficult also insufficient, which degrades generalization performance. relieve dilemma...
In-context learning (ICL) has become the default method for using large language models (LLMs), making exploration of its limitations and understanding underlying causes crucial. In this paper, we find that ICL falls short handling specification-heavy tasks, which are tasks with complicated extensive task specifications, requiring several hours ordinary humans to master, such as traditional information extraction tasks. The performance on these mostly cannot reach half state-of-the-art...
As the main workhorse for model selection, Cross Validation (CV) has achieved an empirical success due to its simplicity and intuitiveness. However, despite ubiquitous role, CV often falls into following notorious dilemmas. On one hand, small data cases, suffers a conservatively biased estimation, since some part of limited hold out validation. other large tends be extremely cumbersome, e.g., intolerant time-consuming, repeated training procedures. Naturally, straightforward ambition is...
Graph neural networks (GNNs) are widely utilized to capture the information spreading patterns in graphs. While remarkable performance has been achieved, there is a new trending topic of evaluating node influence. We propose method influence, which measures prediction change trained GNN model caused by removing node. A real-world application is, "In task predicting Twitter accounts' polarity, had particular account removed, how would others' polarity change?". use as surrogate whose could...
Graph neural networks (GNNs) are widely utilized to capture the information spreading patterns in graphs. While remarkable performance has been achieved, there is a new trending topic of evaluating node influence. We propose method influence, which measures prediction change trained GNN model caused by removing node. A real-world application is, "In task predicting Twitter accounts' polarity, had particular account removed, how would others' polarity change?". use as surrogate whose could...
Low-light image enhancement (LLIE) is vital for autonomous driving. Despite the importance, existing LLIE methods often prioritize robustness in overall brightness adjustment, which can come at expense of detail preservation. To overcome this limitation,we propose Hierarchical Mutual Enhancement via Cross-Attention transformer (ECAFormer), a novel network that utilizes Dual Multi-head Self Attention (DMSA) to enhance both visual and semantic features across scales, significantly preserving...
There have been several recent works proposed to utilize model-based optimization methods improve the productivity of using high-level synthesis (HLS) design domain-specific architectures. They would replace time-consuming performance estimation or simulation with a proxy model, and automatically insert pragmas guide hardware optimizations. In this work, we address challenges associated space exploration (DSE) through evolving landscape HLS tools. As these tools develop, quality results...
High-level synthesis (HLS) is an automated design process that transforms high-level code into hardware designs, enabling the rapid development of accelerators. HLS relies on pragmas, which are directives inserted source to guide process, and pragmas have various settings values significantly impact resulting design. State-of-the-art ML-based methods, such as HARP, first train a deep learning model, typically based graph neural networks (GNNs) applied graph-based representations pragmas....
Few-shot object detection (FSOD) aims to extract semantic knowledge from limited instances of novel categories within a target domain. Recent advances in FSOD focus on fine-tuning the base model based few objects via meta-learning or data augmentation. Despite their success, majority them are grounded with parametric readjustment generalize objects, which face considerable challenges Industry 5.0, such as (i) certain amount time is required, and (ii) parameters constructed being unavailable...
While deep neural networks can achieve good performance on in-distribution samples, their generalization ability significantly degrades under unknown test shifts. We study the problem of out-of-distribution (OOD) capability models by exploring relationship between error and training set size. Previous empirical evidence suggests that falls off as a power size lower errors indicate better model generalization. However, in case OOD this is not true from our observations. Counterintuitively,...
Unsupervised Domain Adaptation (UDA) aims to classify unlabeled target domain by transferring knowledge from labeled source with shift. Most of the existing UDA methods try mitigate adverse impact induced shift via reducing discrepancy. However, such approaches easily suffer a notorious mode collapse issue due lack labels in domain. Naturally, one effective ways this is reliably estimate pseudo for domain, which itself hard. To overcome this, we propose novel method named Progressive...
We argue that the present setting of semisupervised learning on graphs may result in unfair comparisons, due to its potential risk over-tuning hyper-parameters for models. In this paper, we highlight significant influence tuning hyper-parameters, which leverages label information validation set improve performance. To explore limit hyperparameters, propose ValidUtil, an approach fully utilize through extra group hyper-parameters. With even GCN can easily get high accuracy 85.8% Cora. avoid...
We argue that the present setting of semisupervised learning on graphs may result in unfair comparisons, due to its potential risk over-tuning hyper-parameters for models. In this paper, we highlight significant influence tuning hyper-parameters, which leverages label information validation set improve performance. To explore limit hyperparameters, propose ValidUtil, an approach fully utilize through extra group hyper-parameters. With even GCN can easily get high accuracy 85.8% Cora. avoid...
Unsupervised domain adaptation (UDA) aims to transfer knowledge from a well-labeled source different but related unlabeled target with identical label space. Currently, the main workhorse for solving UDA is alignment, which has proven successful. However, it often difficult find an appropriate A more practical scenario so-called partial (PDA) in set or space subsumes one. Unfortunately, PDA, due existence of irrelevant categories domain, quite hard obtain perfect thus resulting mode collapse...