A5101873331- Multimodal Machine Learning Applications
- Topic Modeling
- Genomics and Chromatin Dynamics
- Domain Adaptation and Few-Shot Learning
- RNA Research and Splicing
- Advanced Neural Network Applications
- Mental Health Research Topics
- RNA modifications and cancer
- Advanced Image and Video Retrieval Techniques
- Mobile Crowdsensing and Crowdsourcing
- Evolutionary Algorithms and Applications
- Text and Document Classification Technologies
- Visual Attention and Saliency Detection
- Microtubule and mitosis dynamics
- Data Stream Mining Techniques
- Advanced Text Analysis Techniques
- Cell Adhesion Molecules Research
- Evolution and Genetic Dynamics
- DNA and Nucleic Acid Chemistry
- Cellular Mechanics and Interactions
- Sentiment Analysis and Opinion Mining
- Reinforcement Learning in Robotics
- Expert finding and Q&A systems
- Video Surveillance and Tracking Methods
- Epigenetics and DNA Methylation
Harbin Institute of Technology
2020-2024
University of Pennsylvania
2023-2024
California University of Pennsylvania
2024
Tomorrows Children’s Fund
2023
Communication University of China
2021
Beijing Normal University
2019-2020
Prompt tuning has been employed as an efficient way to adapt large vision-language pre-trained models (e.g. CLIP) various downstream tasks in data-limited or label-limited settings. Nonetheless, visual data (e.g., images) is by default prerequisite for learning prompts existing methods. In this work, we advocate that the effectiveness of image-text contrastive aligning two modalities (for training further makes it feasible treat texts images prompt and introduce TaI prompting. contrast data,...
Abstract In interphase nuclei, chromatin forms dense domains of characteristic sizes, but the influence transcription and histone modifications on domain size is not understood. We present a theoretical model exploring this relationship, considering chromatin-chromatin interactions, modifications, extrusion. predict that heterochromatic governed by balance among diffusive flux methylated histones sustaining them acetylation reactions in process loop extrusion via supercoiling RNAPII at their...
Aiming at the problem of short text classification, this paper proposes a classification method for media data based on RoBERTa and TextRCNN. The semantic vector representation is obtained through used as input TextRCNN training. After experiments THUCNews dataset, RoBERTa-TextRCNN’s accuracy rate reached 94.64%, which 4.53% higher than 0.62% Bert+RCNN. effect better other models, proves its effectiveness in classification.
The interactions between chromatin and the nuclear lamina orchestrate cell type-specific gene activity by forming lamina-associated domains (LADs) which preserve cellular characteristics through repression. However, unlike segments, strength of chromatin-lamina their dependence on environment are not well understood. Here, we develop a theory to predict size shape peripheral heterochromatin considering energetics chromatin-chromatin interactions, affinity kinetics methylation acetylation9in...
With the development of technology, number data is growing rapidly day by day. How to perform efficient big retrieval becomes a critical problem. Similarity-preserving hashing has been widely used in large-scale information because its low storage cost and high computation efficiency. It maps from high-dimensional feature space into binary hamming while preserving similarity. Particularly, deep learning based methods have shown their significantly advantages both effectiveness accuracy....
Parameter-efficient fine-tuning (PEFT) methods have provided an effective way for adapting large vision-language models to specific tasks or scenarios. Typically, they learn a very small scale of parameters pre-trained in white-box formulation, which assumes model architectures be known and accessible. However, are often not open-source due considerations preventing abuse commercial factors, hence posing barrier the deployment PEFT methods. To alleviate dependence on accessibility, we...
Existing video object segmentation (VOS) methods based on matching techniques commonly employ a reference set comprising historical segmented frames, referred to as ‘memory frames’, facilitate the process. However, these suffer from following limitations: (i) Inherent errors in memory frames can propagate and accumulate when utilized templates for subsequent segmentation. (ii) The non-local technique employed top-leading solutions often fails incorporate positional information, potentially...
Policy constraint methods in offline reinforcement learning employ additional regularization techniques to constrain the discrepancy between learned policy and dataset. However, these tend result overly conservative policies that resemble behavior policy, thus limiting their performance. We investigate this limitation attribute it static nature of traditional constraints. In paper, we propose a novel dynamic restricts on samples generated by exponential moving average previously policies. By...
Estimating the quality of answers is one challenges in crowdsourcing. The previous methods focus on estimation for objective tasks, whereas subjective as a common type crowdsourcing have not been well studied. In this paper, we tasks. Considering high uncertainty propose background knowledge enhanced method. More specifically, first learn distributed representation from graphs and text corpora by utilizing multi-task learning framework. Then, construct pseudo-gold answer set each task. Next,...
A model of consciousness proposed by neuroscientists in 1989 is called the theater model, which uses as an analogy to describe "what consciousness". This paper simplifies problem question answering and simulate mechanism human brain. We extract a small amount knowledge from Freebase use it agents' base. Then we build multi-round agent based on Deep Q-learning. train two agents against each other, finally analyze training results. The results show that method designed can dialogue scene well,...
Abstract In interphase nuclei, chromatin is organized into interspersed dense domains with characteristic sizes, both in the nuclear interior and periphery. However, quantitative impact of transcription histone modifications on size distribution these remains unclear. Here, we introduce a mesoscale theoretical model that investigates relationship between heterochromatic domain sizes loop extrusion rates from domains. The considers chromatin-chromatin chromatin-lamina interactions,...
Parameter-efficient fine-tuning (PEFT) methods have provided an effective way for adapting large vision-language models to specific tasks or scenarios. Typically, they learn a very small scale of parameters pre-trained in white-box formulation, which assumes model architectures be known and accessible. However, are often not open-source due considerations preventing abuse commercial factors, hence posing barrier the deployment PEFT methods. To alleviate dependence on accessibility, we...
Prompt tuning has been employed as an efficient way to adapt large vision-language pre-trained models (e.g. CLIP) various downstream tasks in data-limited or label-limited settings. Nonetheless, visual data (e.g., images) is by default prerequisite for learning prompts existing methods. In this work, we advocate that the effectiveness of image-text contrastive aligning two modalities (for training further makes it feasible treat texts images prompt and introduce TaI prompting. contrast data,...