A5088293566- Domain Adaptation and Few-Shot Learning
- Neural Networks and Applications
- Adversarial Robustness in Machine Learning
- Distributed Sensor Networks and Detection Algorithms
- Wireless Communication Security Techniques
- Multimodal Machine Learning Applications
- Privacy-Preserving Technologies in Data
- Machine Learning and ELM
- Machine Learning and Algorithms
- Stochastic Gradient Optimization Techniques
- Anomaly Detection Techniques and Applications
- Sparse and Compressive Sensing Techniques
- Face and Expression Recognition
- Cooperative Communication and Network Coding
- Music and Audio Processing
- Emotion and Mood Recognition
- Advanced Neural Network Applications
- Algorithms and Data Compression
- Advanced Image and Video Retrieval Techniques
- Speech and Audio Processing
- Image Retrieval and Classification Techniques
- Blind Source Separation Techniques
- Error Correcting Code Techniques
- Machine Learning and Data Classification
- Advanced Wireless Communication Techniques
Tsinghua–Berkeley Shenzhen Institute
2017-2025
Tsinghua University
2017-2025
Massachusetts Institute of Technology
2010-2024
University Town of Shenzhen
2024
Tencent (China)
2020
University of California, Los Angeles
2017
National Taiwan University
2009-2016
Tianjin University
2012
As pioneering information technology, the Internet of Things (loT) targets at building an infrastructure embedded devices and networks connected objects, to offer omnipresent ecosystem interaction across billions smart devices, sensors, actuators. The deployment IoT calls for decentralized power supplies, self-powered wireless transmission technologies, which have brought both opportunities challenges existing solutions, especially when network scales up. Triboelectric Nanogenerators...
Abstract Visual understanding on construction sites by deep learning, such as semantic segmentation, is hardly mentioned in the literature due to severe lack of labeled data sets. To resolve this issue, we collect and label 859 images, including 12 classes objects activities, from different sites. We then adopt DeepLabV3+ set with modifications. leverage Cityscape pretrain model, fine‐tune it our collected set. Moreover, multiple augmentation techniques are utilized expand training Our model...
Audio-visual emotion recognition is the research of identifying human emotional states by combining audio modality and visual simultaneously, which plays an important role in intelligent human-machine interactions. With help deep learning, previous works have made great progress for audio-visual recognition. However, these learning methods often require a large amount data training. In reality, acquisition difficult expensive, especially multimodal with different modalities. As result,...
Data heterogeneity is one of the most challenging issues in federated learning, which motivates a variety approaches to learn personalized models for participating clients. One such approach deep neural networks based tasks employing shared feature representation and learning customized classifier head each client. However, previous works do not utilize global knowledge during local also neglect fine-grained collaboration between heads, limit model generalization ability. In this work, we...
Task transfer learning is a popular technique in image processing applications that uses pre-trained models to reduce the supervision cost of related tasks. An important question determine task transferability, i.e. given common input domain, estimating what extent representations learned from source can help target task. Typically, transferability either measured experimentally or inferred through relatedness, which often defined without clear operational meaning. In this paper, we present...
The miniaturization of spectrometer can broaden the application area spectrometry, which has huge academic and industrial value. Among various approaches, filter-based is a promising implementation by utilizing broadband filters with distinct transmission functions. Mathematically, spectral reconstruction be modeled as solving system linear equations. In this paper, we propose an algorithm based on sparse optimization dictionary learning. To verify feasibility algorithm, design implement...
Emotion Recognition in Conversations (ERC) is an increasingly popular task the Natural Language Processing community, which seeks to achieve accurate emotion classifications of utterances expressed by speakers during a conversation. Most existing approaches focus on modeling speaker and contextual information based textual modality, while complementarity multimodal has not been well leveraged, few current methods have sufficiently captured complex correlations mapping relationships across...
We propose two novel transferability metrics fast optimal transport-based conditional entropy (F-OTCE) and joint correspondence OTCE (JC-OTCE) to evaluate how much the source model (task) can benefit learning of target task learn more generalizable representations for cross-domain cross-task transfer learning. Unlike original metric that requires evaluating empirical on auxiliary tasks, our are auxiliary-free such they be computed efficiently. Specifically, F-OTCE estimates by first solving...
One primary focus in multimodal feature extraction is to find the representations of individual modalities that are maximally correlated. As a well-known measure dependence, Hirschfeld-Gebelein-Rényi (HGR) maximal correlation be-´ comes an appealing objective because its operational meaning and desirable properties. However, strict whitening constraints formalized HGR limit application. To address this problem, paper proposes Soft-HGR, novel framework extract informative features from...
Transfer learning across heterogeneous data distributions (a.k.a. domains) and distinct tasks is a more general challenging problem than conventional transfer learning, where either domains or are assumed to be the same. While neural network based feature widely used in applications, finding optimal strategy still requires time-consuming experiments domain knowledge. We propose transferability metric called Optimal Transport Conditional Entropy (OTCE), analytically predict performance for...
With the rapid prevalence and explosive development of Multiplayer Online Battle Arena electronic sports (MOBA esports), much research effort has been devoted to automatically predicting game results (win predictions). While this task great potential in various applications, such as esports live streaming commentator artificial intelligence systems, previous studies fail investigate methods <italic xmlns:mml="http://www.w3.org/1998/Math/MathML"...
Information inequalities govern the ultimate limitations in information theory and as such play an pivotal role characterizing what values entropy of multipartite states can take. Proving inequality, however, quickly becomes arduous when number involved parties increases. For classical systems, [Yeung, IEEE Trans. Inf. Theory (1997)] proposed a framework to prove Shannon-type via linear programming. Here, we derive analogous for quantum based on strong sub-additivity weak monotonicity...
Multi-source transfer learning provides an effective solution to data scarcity in real-world supervised scenarios by leveraging multiple source tasks. In this field, existing works typically use all available samples from sources training, which constrains their training efficiency and may lead suboptimal results. To address this, we propose a theoretical framework that answers the question: what is optimal quantity of needed each task jointly train target model? Specifically, introduce...
The Hirschfeld-Gebelein-Rényi maximal correlation is a well-known measure of statistical dependence between two (possibly categorical) random variables. In inference problems, the functions can be viewed as so called features observed data that carry largest amount information about some latent These are in general non-linear functions, and particularly useful processing high-dimensional data. alternating conditional expectations (ACE) algorithm an efficient way to compute these functions....
Gradient compression (e.g., gradient quantization and sparsification) is a core technique in reducing communication costs distributed learning systems. The recent trend of to use varying number bits across iterations, however, relying on empirical observations or engineering heuristics without systematic treatment analysis. To the best our knowledge, general dynamic that leverages both sparsification techniques still far from understanding. This paper proposes novel Adaptively-Compressed...
Many network information theory problems face the similar difficulty of single letterization. We argue that this is due to lack a geometric structure on space probability distribution. In paper, we develop such by assuming distributions interest are close each other. Under assumption, K-L divergence reduced squared Euclidean metric in an space. Moreover, construct notion coordinate and inner product, which will facilitate solving communication problems. also present application approach...
It is conjectured that the covariance matrices minimizing outage probability under a power constraint for multiple-input multiple-output channels with Gaussian fading are diagonal either zeros or constant values on diagonal. In single-output (MISO) setting, this equivalent to conjecture quadratic forms having largest tail correspond such matrices. This paper provides proof of in MISO setting.
Time series classification is a critical problem in the machine learning field, which spawns numerous research works on it. In this work, we propose AttLSTM-CNNs, an attention-based LSTM network and convolution that jointly extracts underlying pattern among time-series for classification. The automatically captures long-term temporal dependency series, CNN describes spatial sparsity heterogeneity data. extensive experiments show proposed model outperforms other methods
We consider the problem of identifying universal low-dimensional features from high-dimensional data for inference tasks in settings involving learning. For such problems, we introduce natural notions universality and show a local equivalence among them. Our analysis is naturally expressed via information geometry, represents conceptually computationally useful analysis. The development reveals complementary roles singular value decomposition, Hirschfeld-Gebelein-R\'enyi maximal correlation,...
In the time of Big Data, training complex models on large-scale data sets is challenging, making it appealing to reduce volume for saving computation resources by subsampling. Most previous works in subsampling are weighted methods designed help performance subset-model approach full-set-model, hence have no chance acquire a that better than full-set-model. However, we question how can achieve model with less data? this work, propose novel Unweighted Influence Data Subsampling (UIDS) method,...