Yifei Zhang | 张逸飞

I am a first-year Ph.D. student in Computer Science at Carnegie Mellon University, advised by Prof. Max Simchowitz. Previously, I completed my undergraduate studies at the University of Chinese Academy of Sciences (UCAS). I was also a visiting research scholar at the Berkeley AI Research (BAIR) Lab, where I worked with Dr. Qianqian Wang under the guidance of Prof. Angjoo Kanazawa and Prof. Alexei (Alyosha) Efros, and collaborated closely with Prof. Masayoshi Tomizuka. My prior experience also includes an internship at the Institute for AI Industry Research (AIR), Tsinghua University, where I was advised by Prof. Hao Zhao.

Email  /  CV  /  Scholar  /  Github

Research

I'm interested in building learning systems that are reliable in the real world—systems that can continuously learn from experience and adapt as the world changes. More concretely, I'm drawn to problems in representation learning, online adaptation, and active learning, motivated by a simple question: can we enable models to learn effectively from ongoing experience, in a way that resembles how humans learn? Selected works are listed below (*equal contribution).

Continuous 3D Perception Model with Persistent State

A new framework for reasoning about the 3D world in an online, sequential manner. Given an input image stream, our method simultaneously updates an inner state with the current observation and reads from it to make predictions of 3D geometry & camera pose for the current view, as well as infer unseen portions of the scene.

Sparse Diffusion Policy: A Sparse, Reusable, and Flexible Policy for Robot Learning

We propose a Sparse Diffusion Policy (SDP) that integrates a Mixture of Experts module specifically designed for multitask learning, continual learning and rapid adaptation to new tasks.

FastMAC: Stochastic Spectral Sampling of Correspondence Graph

We propose a new technique of stochastic spectral sampling of correspondence graph and build a complete 3D registration pipeline that reaches real-time speed while leading to little to none performance drop.

Dual-frame Fluid Motion Estimation with Test-time Optimization and Zero-divergence Loss

A new fluid motion tracking method that is completely self-supervised and notably outperforms its supervised counterparts while requiring only 1% of the training samples (without labels) used by previous methods.

Selected Honors and Awards

Designed based on Jon Barron's website