Tianyi Chen 陈天翼

I am currently a Machine Learning PhD student at Georgia Tech, co-advised by Prof. Kai Wang and Prof. Bo Dai. Prior to my PhD, I was an undergraduate at Shanghai Jiao Tong University (SJTU), where I majored in Artificial Intelligence.

During my undergraduate studies, I had the opportunity to be a visiting student at the National University of Singapore (NUS). There, I worked under the guidance of Prof. Lin Shao.

In addition, I have worked as a Research Assistant at the MOE Key Lab of Artificial Intelligence, supervised by Prof. Liqing Zhang and Prof. Jianfu Zhang. Previously, I also had the chance to collaborate with Prof. Junchi Yan at the SJTU-ReThinkLab.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research

My research interests mainly lie in reinforcement learning and optimization for social impact, I also have experience in robotics, computer vision and differentiable constraints.

Grasping Points Discovery for Cloth Manipulation via Differentiable Physics-based Simulation

A novel grasping point optimization scheme that leverages gradient information from the differentiable simulation process. We Conducted extensive experiments of clothes grasping tasks on manipulating hats, socks and bags and our gradient-based optimization scheme successfully optimizes grasp points to achieve optimal results.

UniInsertion: A Unified Model-based Insertion Skill Learning via Differentiable Physics-based Simulation

Presented a model-based framework leveraging differentiable physics simulation and the innovative concept of “learning from reversal” to enable robotic insertion of both rigid and deformable objects.

A Novel Evaluation Framework for Image Inpainting via Multi-Pass Self-Consistency

A novel image inpainting evaluation framework that leverages the power of aggregated multi-pass image inpainting. Our self-supervised metric performs in scenarios with or without unmasked images and mitigates biases toward specific inpainting solutions.

LinSATNet: The Positive Linear Satisfiability Neural Networks

Proposed the first differentiable satisfiability layer based on an extension of the classic Sinkhorn algorithm for jointly encoding multiple sets of marginal distributions. We showcase our technique in solving constrained (specifically satisfiability) problems by one-shot neural networks, including i) a neural routing solver learned without supervision of optimal solutions; ii) a partial graph matching network handling graphs with unmatchable outliers on both sides; iii) a predictive network for financial portfolios with continuous constraints.

Projects

These include coursework, side projects and unpublished research work.

Design and source code from Jon Barron's website