A5101990494- Reinforcement Learning in Robotics
- Domain Adaptation and Few-Shot Learning
- Human Pose and Action Recognition
- Topic Modeling
- Machine Learning and Data Classification
- Machine Learning and Algorithms
- Advanced Neural Network Applications
- Advanced Bandit Algorithms Research
- Face and Expression Recognition
- Multimodal Machine Learning Applications
- Advanced Vision and Imaging
- Computer Graphics and Visualization Techniques
- Neural Networks and Applications
- 3D Shape Modeling and Analysis
- Stochastic Gradient Optimization Techniques
- Video Surveillance and Tracking Methods
- Generative Adversarial Networks and Image Synthesis
- Sparse and Compressive Sensing Techniques
- Anomaly Detection Techniques and Applications
- Natural Language Processing Techniques
- Text and Document Classification Technologies
- Gaussian Processes and Bayesian Inference
- Advanced Graph Neural Networks
- Machine Learning and ELM
- Adversarial Robustness in Machine Learning
Shanghai Artificial Intelligence Laboratory
2022-2024
Shanghai Jiao Tong University
2024
ShangHai JiAi Genetics & IVF Institute
2022-2024
University of Electronic Science and Technology of China
2006-2023
Nanyang Technological University
2021-2022
Google (United States)
2019-2022
Tsinghua University
1999-2020
State Key Laboratory of Mobile Networks and Mobile Multimedia Technology
2020
Brain (Germany)
2019-2020
Alphabet (United States)
2018-2020
Human skeleton, as a compact representation of human action, has received increasing attention in recent years. Many skeleton-based action recognition methods adopt GCNs to extract features on top skeletons. Despite the positive results shown these attempts, GCN-based are subject limitations robustness, interoperability, and scalability. In this work, we propose PoseConv3D, new approach recognition. PoseConv3D relies 3D heatmap volume instead graph sequence base Compared methods, is more...
Kernel classifiers and regressors designed for structured data, such as sequences, trees graphs, have significantly advanced a number of interdisciplinary areas computational biology drug design. Typically, kernels are beforehand data type which either exploit statistics the structures or make use probabilistic generative models, then discriminative classifier is learned based on via convex optimization. However, an elegant two-stage approach also limited kernel methods from scaling up to...
Deep generative models have been enjoying success in modeling continuous data. However it remains challenging to capture the representations for discrete structures with formal grammars and semantics, e.g., computer programs molecular structures. How generate both syntactically semantically correct data still largely an open problem. Inspired by theory of compiler where syntax semantics check is done via syntax-directed translation (SDT), we propose a novel variational autoencoder (SD-VAE)...
When function approximation is used, solving the Bellman optimality equation with stability guarantees has remained a major open problem in reinforcement learning for decades. The fundamental difficulty that operator may become an expansion general, resulting oscillating and even divergent behavior of popular algorithms like Q-learning. In this paper, we revisit equation, reformulate it into novel primal-dual optimization using Nesterov's smoothing technique Legendre-Fenchel transformation....
Neural networks are a powerful class of nonlinear functions that can be trained end-to-end on various applications. While the over-parametrization nature in many neural renders ability to fit complex and strong representation power handle challenging tasks, it also leads highly correlated neurons hurt generalization incur unnecessary computation cost. As result, how regularize network avoid undesired redundancy becomes an important issue. To this end, we draw inspiration from well-known...
The general perception is that kernel methods are not scalable, so neural nets become the choice for large-scale nonlinear learning problems. Have we tried hard enough methods? In this paper, propose an approach scales up using a novel concept called doubly stochastic functional gradients. Based on fact many can be expressed as convex optimization problems, our solves problems by making two unbiased approximations to gradient—one random training points and another features associated with...
Retrosynthesis is one of the fundamental problems in organic chemistry. The task to identify reactants that can be used synthesize a specified product molecule. Recently, computer-aided retrosynthesis finding renewed interest from both chemistry and computer science communities. Most existing approaches rely on template-based models define subgraph matching rules, but whether or not chemical reaction proceed defined by hard decision rules. In this work, we propose new approach using...
Semi-supervised action recognition is a challenging but important task due to the high cost of data annotation. A common approach this problem assign unlabeled with pseudo-labels, which are then used as additional supervision in training. Typically recent work, pseudo-labels obtained by training model on labeled data, and using confident predictions from teach itself. In we propose more effective pseudo-labeling scheme, called Cross-Model Pseudo-Labeling (CMPL). Concretely, introduce...
The general perception is that kernel methods are not scalable, and neural nets the of choice for nonlinear learning problems. Or have we simply tried hard enough methods? Here propose an approach scales up using a novel concept called "doubly stochastic functional gradients". Our relies on fact many can be expressed as convex optimization problems, solve problems by making two unbiased approximations to gradient, one random training points another functions associated with kernel, then...
Learning-based binary hashing has become a powerful paradigm for fast search and retrieval in massive databases. However, due to the requirement of discrete outputs hash functions, learning such functions is known be very challenging. In addition, objective adopted by existing techniques are mostly chosen heuristically. this paper, we propose novel generative approach learn through Minimum Description Length principle that learned codes maximally compress dataset can also used regenerate...
Many machine learning tasks, such as with invariance and policy evaluation in reinforcement learning, can be characterized problems of from conditional distributions. In problems, each sample $x$ itself is associated a distribution $p(z|x)$ represented by samples $\{z_i\}_{i=1}^M$, the goal to learn function $f$ that links these distributions target values $y$. These become very challenging when we only have limited or extreme case one distribution. Commonly used approaches either assume $z$...
Existing approaches for named entity recognition suffer from data sparsity problems when conducted on short and informal texts, especially user-generated social media content. Semantic augmentation is a potential way to alleviate this problem. Given that rich semantic information implicitly preserved in pre-trained word embeddings, they are ideal resources augmentation. In paper, we propose neural-based approach NER texts where both local (from running text) augmented semantics taken into...
Digital human motion synthesis is a vibrant research field with applications in movies, AR/VR, and video games. Whereas methods were proposed to generate natural realistic motions, most only focus on modeling humans largely ignore object movements. Generating task-oriented human-object interaction motions simulation challenging. For different intents of using the objects, conduct various which requires first approach objects then make them move consistently instead staying still. Also,...
In this paper, we consider the problem of machine teaching, inverse learning. Different from traditional teaching which views learners as batch algorithms, study a new paradigm where learner uses an iterative algorithm and teacher can feed examples sequentially intelligently based on current performance learner. We show that complexity in case is very different case. Instead constructing minimal training set for learners, our focuses achieving fast convergence model. Depending level...
An important problem that arises in reinforcement learning and Monte Carlo methods is estimating quantities defined by the stationary distribution of a Markov chain. In many real-world applications, access to underlying transition operator limited fixed set data has already been collected, without additional interaction with environment being available. We show consistent estimation remains possible this challenging scenario, effective can still be achieved applications. Our approach based...
With wider application of deep neural networks (DNNs) in various algorithms and frameworks, security threats have become one the concerns. Adversarial attacks disturb DNN-based image classifiers, which attackers can intentionally add imperceptible adversarial perturbations on input images to fool classifiers. In this paper, we propose a novel purification approach, referred as guided diffusion model for (GDMP), help protect classifiers from attacks. The core our approach is embed into...
Neural radiance fields (NeRF) and its subsequent variants have led to remarkable progress in neural rendering. While most of recent rendering works focus on objects small-scale scenes, developing methods for city-scale scenes is great potential many real-world applications. However, this line research impeded by the absence a comprehensive high-quality dataset, yet collecting such dataset over real costly, sensitive, technically infeasible. To end, we build large-scale, comprehensive,...
Synthetic data has emerged as a promising source for 3D human research it offers low-cost access to large-scale datasets. To advance the diversity and annotation quality of models, we introduce new synthetic dataset, SynBody, with three appealing features: 1) clothed parametric model that can generate diverse range subjects; 2) layered representation naturally high-quality annotations support multiple tasks; 3) scalable system producing realistic facilitate real-world tasks. The dataset...
A goal of artificial intelligence is to construct an agent that can solve a wide variety tasks. Recent progress in text-guided image synthesis has yielded models with impressive ability generate complex novel images, exhibiting combinatorial generalization across domains. Motivated by this success, we investigate whether such tools be used more general-purpose agents. Specifically, cast the sequential decision making problem as text-conditioned video generation problem, where, given...