Benlin Liu

I am a Ph.D. student in Paul G. Allen School of Computer Science & Engineering, University of Washington, working with Ranjay Krishna. I am affiliated with UW Graphics and Imaging Laboratory (GRAIL).

I receive my master degree from Department of Computer Science, UCLA, where I was a research assistant under the supervision of Prof. Cho-jui Hsieh . I also collaborated with Prof. Xiaolong Wang at UCSD . I obtained my BEng. degree from the Department of Electronic Engineering , Tsinghua University , and I worked with Prof. Jiwen Lu of the Department of Automation . During undergraduate, I visited the GRASP Lab at University of Pennsylvania and worked with Prof. Jianbo Shi.

My current research interest is at the intersection of compution vision and machine learning, with a current focus on diffusion, self-supervised learning and large language model. Past research is more about efficient machine learning model and how to learn more generlizable vision model in a data-efficient way.

Email  /  CV  /  Google Scholar  /  Github  /  Twitter

* indicates equal contribution

Enhancing large language model efficiency, especially in multimodal contexts, with 'importance-driven cache merging' to manage KV cache memory needs, thereby boosting long instruction following and long output generation in multimodal chatbots!

Fine-grained and accurate evaluation of synthesized images using Image-to-Text Models (e.g. GPT-4, BLIP-2, etc.) and Large Language Models (e.g. GPT-3.5). More accurate than CLIP!

Text-to-Image generation models (e.g. Stable Diffusion) are not only for creating cool stuff, but can also be applied to multiple dense prediction tasks!

We propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically for vision transformer acceleration.

We propose to generate random layouts of a scene by making use of the objects in the synthetic CAD dataset and learn the 3D scene representation by applying object-level contrastive learning on two random scenes generated from the same set of synthetic objects.

We introduce a new self-supervised task on videos to improve the out-of-domain generalization of object detectors.

We present a new Multi-Proxy Wasserstein Classifier to imporve the image classification models by calculating a non-uniform matching flow between the elements in the feature map of a sample and multiple proxies of a class using optimal transport theory.

We boost the performance of CNNs by learning soft targets for shallow layers via meta-learning.

Website Template