A5101848358- Natural Language Processing Techniques
- X-ray Diffraction in Crystallography
- Crystallization and Solubility Studies
- Topic Modeling
- Neural Networks and Applications
- Neural dynamics and brain function
- Music Technology and Sound Studies
- Neural Networks and Reservoir Computing
- Music and Audio Processing
- Sentiment Analysis and Opinion Mining
- Reinforcement Learning in Robotics
- Stochastic Gradient Optimization Techniques
- Remote Sensing and Land Use
- Interpreting and Communication in Healthcare
- Clay minerals and soil interactions
- Speech and Audio Processing
- Iron oxide chemistry and applications
- Generative Adversarial Networks and Image Synthesis
- Land Use and Ecosystem Services
- Human Pose and Action Recognition
- Multimodal Machine Learning Applications
- Speech and dialogue systems
- Domain Adaptation and Few-Shot Learning
- Image and Video Quality Assessment
- Motor Control and Adaptation
Peking University
2017-2023
Jinan University
2023
Xihua University
2022
China University of Petroleum, Beijing
2021
Chongqing University of Posts and Telecommunications
2021
Tianjin University
2011-2020
Friedrich Miescher Institute
2020
Renmin University of China
2020
Constructor University
2018-2019
University of Pennsylvania
2019
Virtual reality (VR), a new type of simulation and interaction technology, has aroused widespread attention research interest. It is necessary to evaluate the VR quality provide standard for rapidly developing technology. To best our knowledge, few researchers have built benchmark databases designed related algorithms, which hindered further development In this paper, free available data set (VRQ-TJU) assessment proposed with subjective scores each sample data. The validity database been...
Tianlin Liu, Lyle Ungar, João Sedoc. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.
Graph convolutional networks (GCN) have attracted increasing interest in action recognition recent years. GCN models human skeleton sequences as spatio-temporal graphs. Also, attention mechanisms are often jointly used with GCNs to highlight important frames or body joints a sequence. However, modules learn parameters offline and fixed, so may not adapt well unseen samples. In this paper, we propose simple but effective motion-driven spatial temporal adaptation strategy dynamically...
Energy-based models (EBMs) offer a flexible framework for parameterizing probability distributions using neural networks. However, learning EBMs by exact maximum likelihood estimation (MLE) is generally intractable, due to the need compute partition function (normalization constant). In this paper, we propose novel formulation approximately probabilistic in combinatorially-large discrete spaces, such as sets or permutations. Our key idea jointly learn both an energy model and its...
Word vectors are at the core of many natural language processing tasks. Recently, there has been interest in post-processing word to enrich their semantic information. In this paper, we introduce a novel vector technique based on matrix conceptors (Jaeger 2014), family regularized identity maps. More concretely, propose use suppress those latent features having high variances. The proposed method is purely unsupervised: it does not rely any corpus or external linguistic database. We evaluate...
Panoramic video and stereoscopic panoramic are essential carriers of virtual reality content, so it is very crucial to establish their quality assessment models for the standardization industry. However, challenging evaluate at present. One reason that spatial information warped due projection process, conventional (VQA) method difficult deal with this problem. Another traditional VQA problematic capture complex global time in video. In response above questions, paper presents an end-to-end...
Mixture-of-Experts (MoE) models are a promising way to scale up model capacity without significantly increasing computational cost. A key component of MoEs is the router, which decides subset parameters (experts) process feature embeddings (tokens). In this paper, we present comprehensive study routers in for computer vision tasks. We introduce unified MoE formulation that subsumes different with two parametric routing tensors. This covers both sparse MoE, uses binary or hard assignment...
In many Reinforcement Learning (RL) papers, learning curves are useful indicators to measure the effectiveness of RL algorithms. However, complete raw data rarely available. As a result, it is usually necessary reproduce experiments from scratch, which can be time-consuming and error-prone. We present Open Benchmark, set fully tracked experiments, including not only usual such as episodic return, but also all algorithm-specific system metrics. Benchmark community-driven: anyone download,...
Analog, unclocked, spiking neuromorphic microchips open new perspectives for implantable or wearable biosensors and biocontrollers, due to their low energy consumption heat dissipation. However, the challenges from a computational point of view are formidable. Here we outline our solutions realize reservoir computing paradigm on such hardware address combined problems bit resolution, device mismatch, approximate neuron models, timescale mismatch. The main contribution is scheme, called...
Deep neural networks have dramatically transformed machine learning, but their memory and energy demands are substantial. The requirements of real biological rather modest in comparison, one feature that might underlie this austerity is sparse connectivity. In deep trainable perform well on a specific task usually constructed using label-dependent pruning criteria. article, we introduce Neural Tangent Transfer, method instead finds label-free manner. Specifically, find whose training...
Distributed representations of words, better known as word embeddings, have become important building blocks for natural language processing tasks. Numerous studies are devoted to transferring the success unsupervised embeddings sentence embeddings. In this paper, we introduce a simple representation sentences in which embedding is represented weighted average vectors followed by soft projection. We demonstrate effectiveness proposed method on clinical semantic textual similarity task...
Approximately 100 Intellectual Property Information Services Centres have been established in Chinese university libraries, more than 80% of them since 2017. The context this boom is the rapidly increasing number patent applications China, as well an unacceptably low transfer ratio. Do represent a promising direction for library transformation? This central issue addressed article. characteristics evolutionary path and driving forces are discussed, distinctive intellectual property...
Distributional representations of words, also known as word vectors, have become crucial for modern natural language processing tasks due to their wide applications. Recently, a growing body vector postprocessing algorithm has emerged, aiming render off-the-shelf vectors even stronger. In line with these investigations, we introduce novel scheme under causal inference framework. Concretely, the pipeline is realized by Half-Sibling Regression (HSR), which allows us identify and remove...
Distributed representations of sentences have become ubiquitous in natural language processing tasks. In this paper, we consider a continual learning scenario for sentence representations: Given sequence corpora, aim to optimize the encoder with respect new corpus while maintaining its accuracy on old corpora. To address problem, propose initialize encoders help corpus-independent features, and then sequentially update using Boolean operations conceptor matrices learn corpus-dependent...
Regularized optimal transport (OT) is now increasingly used as a loss or matching layer in neural networks. Entropy-regularized OT can be computed using the Sinkhorn algorithm but it leads to fully-dense transportation plans, meaning that all sources are (fractionally) matched with targets. To address this issue, several works have investigated quadratic regularization instead. This preserves sparsity and unconstrained smooth (semi) dual objectives, solved off-the-shelf gradient methods....
Aligning language models with human preferences is crucial for reducing errors and biases in these models. Alignment techniques, such as reinforcement learning from feedback (RLHF), are typically cast optimizing a tradeoff between preference rewards proximity regularization term that encourages staying close to the unaligned model. Selecting an appropriate level of critical: insufficient can lead reduced model capabilities due reward hacking, whereas excessive hinders alignment. Traditional...
Direct alignment from preferences (DAP) methods, such as DPO, have recently emerged efficient alternatives to reinforcement learning human feedback (RLHF), that do not require a separate reward model. However, the preference datasets used in DAP methods are usually collected ahead of training and never updated, thus is purely offline. Moreover, responses these often sampled language model distinct one being aligned, since evolves over training, phase inevitably off-policy. In this study, we...