Hongyi Zhang
Hongyi Zhang profile photo

Hongyi Zhang

I am a senior scientist at PokeBot, working towards AGI in the physical world. I spent 7 years at ByteDance, where I built a production-level browser-use agent framework, trained LLMs, and invented an MPC-based online bidding system to power ByteDance's multi-billion dollar ad business.

I obtained my Ph.D. in Brain and Cognitive Sciences (BCS) and Laboratory for Information and Decision Systems (LIDS) at MIT advised by Prof. Suvrit Sra. I got my B.S. in Machine Intelligence from Peking University, advised by Prof. Liwei Wang. I had three research internships, two at Facebook AI Research (FAIR) and one at Toyota Technological Institute at Chicago (TTIC).

See below for my representative research projects.

Updates

Embodied AI Research

I have joined PokeBot Inc. as a senior scientist. Next plan: Physical AGI :)

Project Highlights

Mixup: Beyond Empirical Risk Minimization

Mixup: Beyond Empirical Risk Minimization

Hongyi Zhang, Moustafa Cisse, Yann N. Dauphin, David Lopez-Paz

data augmentation generalization

ICLR 2018

Mixup is a simple yet effective data augmentation technique that improves the generalization and robustness of machine learning models by blending together multiple training examples. It is the most cited data augmentation technique in machine learning.

Paper Code
Fixup Initialization: Residual Learning Without Normalization

Fixup Initialization: Residual Learning Without Normalization

Hongyi Zhang, Yann N. Dauphin, Tengyu Ma

deep learning optimization

ICLR 2019

We prove that normalization modules stabilize the training of residual networks because they avoid the exploding gradient problem in vanilla initialization, and propose a new initialization method that allows for stable and efficient residual learning without normalization.

Paper Code
First-order Methods for Geodesically Convex Optimization

First-order Methods for Geodesically Convex Optimization

Hongyi Zhang, Suvrit Sra

optimization theory

COLT 2016

We derive a novel trigonometric inequality for Riemannian manifolds, which is a generalization of "law of cosines" in Euclidean space, and is crucial for the analysis of Riemannian optimization algorithms. The inequality enables us to establish the first-ever iteration complexity bounds for first-order optimization methods on geodesically convex functions.

Paper