A5023501811- Topic Modeling
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Domain Adaptation and Few-Shot Learning
- Machine Learning and Data Classification
- Bayesian Methods and Mixture Models
- Advanced Image and Video Retrieval Techniques
- Speech Recognition and Synthesis
- Caching and Content Delivery
- Text and Document Classification Technologies
- Adversarial Robustness in Machine Learning
- Complex Network Analysis Techniques
- Gaussian Processes and Bayesian Inference
- Advanced Computational Techniques and Applications
- Advanced Sensor and Control Systems
- Data Quality and Management
- Parallel Computing and Optimization Techniques
- Vehicular Ad Hoc Networks (VANETs)
- Distributed and Parallel Computing Systems
- Cancer-related molecular mechanisms research
- Blockchain Technology Applications and Security
- Music and Audio Processing
- Advanced Graph Neural Networks
- Network Security and Intrusion Detection
- Image Retrieval and Classification Techniques
Xinjiang Normal University
2022-2024
China National Petroleum Corporation (China)
2024
Qingdao University
2024
Google (United States)
2014-2023
Nanjing University of Information Science and Technology
2023
Beihang University
2012-2022
Dalian University of Technology
2021-2022
Lawrence Berkeley National Laboratory
2021
Chongqing University of Technology
2021
Harvard University Press
2020
We present a new dataset of image caption annotations, Conceptual Captions, which contains an order magnitude more images than the MS-COCO (Lin et al., 2014) and represents wider variety both styles. achieve this by extracting filtering annotations from billions webpages. also quantitative evaluations number captioning models show that model architecture based on Inception-ResNetv2 (Szegedy 2016) for image-feature extraction Transformer (Vaswani 2017) sequence modeling achieves best...
The availability of large-scale image captioning and visual question answering datasets has contributed significantly to recent successes in vision-and-language pretraining. However, these are often collected with overrestrictive requirements inherited from their original target tasks (e.g., caption generation), which limit the resulting dataset scale diversity. We take a step further pushing limits pretraining data by relaxing collection pipeline used Conceptual Captions 3M (CC3M) [54]...
Sharan Narang, Hyung Won Chung, Yi Tay, Liam Fedus, Thibault Fevry, Michael Matena, Karishma Malkan, Noah Fiedel, Noam Shazeer, Zhenzhong Lan, Yanqi Zhou, Wei Li, Nan Ding, Jake Marcus, Adam Roberts, Colin Raffel. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.
Soravit Changpinyo, Doron Kukliansy, Idan Szpektor, Xi Chen, Nan Ding, Radu Soricut. Proceedings of the 2022 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2022.
Ye Zhang, Nan Ding, Radu Soricut. Proceedings of the 2018 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.
We introduce "talking-heads attention" - a variation on multi-head attention which includes linearprojections across the attention-heads dimension, immediately before and after softmax operation.While inserting only small number of additional parameters moderate amount additionalcomputation, talking-heads leads to better perplexities masked language modeling tasks, aswell as quality when transfer-learning comprehension question answering tasks.
In view of most current studies on text sentiment classification focus the deep learning model to obtain sentimental characteristics English text. Chinese analysis is rarely involved, and only context information statement considered, but syntax considered. this paper, a novel proposed (Dependency Tree Graph Convolutional Network, DTGCN) combined syntactically dependent tree with graph convolution. Firstly, Bi-GRU (Bi-directional Gated Recurrent Unit) used learn contextual feature...
We consider classification problems in which the label space has structure. A common example is hierarchical spaces, corresponding to case where one subsumes another (e.g., animal dog). But labels can also be mutually exclusive dog vs cat) or unrelated furry, carnivore). To jointly model hierarchy and exclusion relations, notion of a HEX (hierarchy exclusion) graph was introduced [8]. This combined conditional random field (CRF) with deep neural network (DNN), resulting state art results...
In applications we may want to compare different document collections: they could have shared content but also and unique aspects in particular collections. This task has been called comparative text mining or cross-collection modeling. We present a differential topic model for this application that models both differences similarities. For use hierarchical Bayesian nonparametric models. Moreover, found it was important properly power-law phenomena topic-word distributions thus used the full...
Policy-gradient approaches to reinforcement learning have two common and undesirable overhead procedures, namely warm-start training sample variance reduction. In this paper, we describe a method based on softmax value function that requires neither of these procedures. Our combines the advantages policy-gradient methods with efficiency simplicity maximum-likelihood approaches. We apply new cold-start in sequence generation models for structured output prediction problems. Empirical evidence...
The Hidden Markov Model (HMM) has been widely used in many applications such as speech recognition. A common challenge for applying the classical HMM is to determine structure of hidden state space. Based on Dirichlet Process, a nonparametric Bayesian proposed, which allows an infinite number states and uses Gaussian components support continuous observations. An efficient variational inference method also proposed applied model. Our experiments demonstrate that new model can discover both...
Sequence generation models trained with teacher-forcing suffer from issues related to exposure bias and lack of differentiability across timesteps. Our proposed method, Teacher-Forcing N-grams (TeaForN), addresses both these problems directly, through the use a stack N decoders decode along secondary time axis that allows model-parameter updates based on prediction steps. TeaForN can be used wide class decoder architectures requires minimal modifications standard setup. Empirically, we show...
We develop dependent hierarchical normalized random measures and apply them to dynamic topic modeling. The dependency arises via superposition, subsampling point transition on the underlying Poisson processes of these measures. used include normalised generalised Gamma that demonstrate power law properties, unlike Dirichlet previously in Inference for model includes adapting a recently developed slice sampler directly manipulate process. Experiments performed news, blogs, academic Twitter...
Despite recent advances in its theoretical understanding, there still remains a significant gap the ability of existing PAC-Bayesian theories on meta-learning to explain performance improvements few-shot learning setting, where number training examples target tasks is severely limited. This originates from an assumption which supposes that observed and follow same distribution, rarely holds practice. By relaxing this assumption, we develop two bounds tailored for setting show algorithms...
Geometrical approaches used in Marcum Q-Function have been successfully applied former works to find its bounds. However, this bound is still not tight enough for mainly two reasons: first, those approximations contain only different radiuses of arcs; second, the approximation restricted at form circle whose radius a constant, while varies continuously with direction. In paper, we main aims: want deeper analytical reasoning on why geometrical way works; try solve drawbacks and propose new...
Blockchain stands out in addressing the data security requirements of Internet Vehicles. However, blockchain has storage pressure that cannot be met by most existing nodes. The emergence Mobile Edge Computing allows nodes closer to users undertake caching and computation process. Although sharding can alleviate on nodes, frequent cross-shard communication affect overall performance blockchain. In this paper, combining features traffic flow with strong regional similarity as well inter-node...
We introduce a new multi-modal task for computer systems, posed as combined vision-language comprehension challenge: identifying the most suitable text describing scene, given several similar options. Accomplishing entails demonstrating beyond just recognizing "keywords" (or key-phrases) and their corresponding visual concepts. Instead, it requires an alignment between representations of two modalities that achieves visually-grounded "understanding" various linguistic elements dependencies....
This paper introduces a new kind of hybrid Cache coherence protocol-MECSIF, which applicants for multiprocessor environment, based on cache line write strategy. Through the introduction small dictionary-D-Cache in system architecture, protocol overcomes shortcoming snoopy that data request was undifferentiated broadcasted. Protocol extends block state so eliminates "ping-pang" phenomenon, uses strategy to reduce L1 miss ratio. Simulation results show MECSIF extent improves efficiency...