Research
Current Research Projects
With the emergence of larger and larger neural model architectures and new learning tasks, it becomes harder to deploy these powerful models into real-world applications, and to understand their behavior and potential shortcomings. My research aims to explore the interplay between model architecture, training data, and loss and regularization landscape of neural network models for better efficiency and robustness, while exploring new learning schemes that codesigns learning task (data) and model architecture to achieve more controllable and interpretable deep learning. Here are some ongoing projects of my research:
Compressing emerging neural network architectures and novel learning tasks
Post-training quantization for vision transformer, diffusion model, BEVFormer, etc., with potential extension onto LLM and multi-modal fundation models
Finegrained understanding and improvement of (large) model performance
Understanding the impact of model compression on the loss landscape of neural network models
Evaluating and improving model performance on specific subtask with regularized optimization and active finetuning
Systematically evaluate and actively resolve the performance gap of compressed large fundation models
Learning diverse ensemble model for better generalizability, robustness, and interpretability
Ensemble of heterogeneously compressed model
Diverse ensemble training with loss landscape characteristics
Codesign learning task and architecture for ensemble submodels
Past Research Projects
I have been working on improving the efficiency and robustness of deep neural networks during my PhD study at Duke University. You can find a brief summary of my previous research projects here.
Besides my main research topics, I also participated in the research of privacy preservation, federated learning, SW/HW codesign of emerging devices architectures, and neural network accelerator architecture design. You can check out my full research specturm in my publication list.
Research Fundings
My research are partially funded by the following companies and grants, where I serve as the key researcher in these projects.
Panasonic through BAIR Open Research Commons
Defense Advanced Research Projects Agency (DARPA) Grant – HR00111990079
Robust Ensemble Generation from Distilled Feature Transforms (REG-DFT)
Defense Advanced Research Projects Agency (DARPA) Grant – HR00112090054
Neural-Network Enhanced Radar Surveillance (NNERS)
National Science Foundation (NSF) Grant – 2112562
AI Institute for Edge Computing Leveraging Next Generation Networks (Athena)
National Science Foundation (NSF) Grant – 1822085
Industry-University Cooperative Research Center (IUCRC) for Alternative Sustainable and Intelligent Computing (ASIC)