Yatong Bai

Yatong Bai 

Ph.D. Candidate

University of California, Berkeley

Advisor: Somayeh Sojoudi

Email: yatong_bai (at) berkeley.edu

Research Portfolio Slides

I am looking for full-time research/engineering opportunities that start in 2025.

My Journey

I am a fifth-year Ph.D. candidate at the mechanical engineering department of UC Berkeley, advised by Professor Somayeh Sojoudi.

Prior to joining Berkeley, I obtained Bachelor's degrees in computer engineering and mechanical engineering from Georgia Institute of Technology, where I researched with Professors Julien Meaud and Thomas Conte.

I became increasingly interested in machine learning in my senior year, and decided to pursue a Ph.D. in the field, focusing on efficient and reliable deep learning and generative AI.

I have interned at Adobe Research, Microsoft, Scale AI, Honda Aircraft Company, and Tesla.

Research Interests

My research aims to make deep learning and generative AI more efficient, aligned, robust, and reliable. My interests span across generative models (particulaly audio/music), robust deep learning, (convex) optimization, reinforcement learning, and controls.

Specifically, my research areas so far include:

  • Diffusion Models (Audio/Music Generation).

    • Accelerating diffusion-based text-to-audio generation with consistency distillation (ConsistencyTTA).
    • Using reinforcement learning to align text-to-music generation to human preferences (ongoing).
  • Robust Deep Learning.

    • Analyzing the vulnerabilities of large language models (LLMs) when used in conversational search engines (RAGDOLL).
    • Mixing classifiers to address the "accuracy-robustness trade-off" for image classifiers (MixedNUTS, Adaptive Smoothing).
  • Convex Neural Network Optimization.

    • Efficient algorithms for training two-layer ReLU neural networks with global optimality guarantees (link).
    • Convex formulations for adversarial training, which build two-layer networks with adversarial robustness (link).

News