Ziyuan Zhong

I am currently a Quantitative Researcher at Squarepoint. Previously, I finished my Ph.D. at the Department of Computer Science at Columbia University. I am fortunate to be advised by Prof. Baishakhi Ray. I did my undergrads at Reed College and Columbia University.

Email  /  Google Scholar  /  Github

My PhD research work mainly focuses on testing/improving robustness of learning based systems like Autonomous Driving Systems (ADSs) and Image Classifiers. Representative papers are highlighted.

In this dissertation, we first identified the neceesity of efficient simulation testing for Autonomous Driving Systems (ADSs) lies in crafting optimal testing scenarios. Next, we proposed four methods (AutoFuzz, FusED, CTG, and CTG++) to efficiently construct these testing scenarios.

We propose a method that sidesteps the limitations of existing work on system performance optimization under environmental changes by identifying invariant causal predictors under environmental changes, enabling the optimization process to operate on a reduced search space, leading to faster system performance optimization.

We leveraged Large Langauge Models (LLMs) and developed a language guided scene-level conditional diffusion model for language-controllable, realistic traffic generation.

We developed a conditional diffusion model (trained on large-scale traffic flow datasets) that can generate controllable, feasible, and realistic traffic trajectories.

We propose a method that can automatically generate a more concise map from a given complex map to reduce redundant test cases.

A series of methods based on weighted regularization for repairing group-level errors of DNNs.

FusED can efficiently identify fusion errors.

A Survey on Scenario-Based Testing for Automated Driving Systems in High-Fidelity Simulation.

AutoFuzz uses a grammar-based, learning-guided fuzzing technique to efficiently find violations of Autonomous Driving Systems.

DeepRobust can identify the input images whose small variations may lead to erroneous DNN behaviors.

We developed a testing technique, DeepInspect, to automatically detect class-based confusion and bias errors in DNN-driven image classification software.

We propose to regularize the representation space under adversarial attack with metric learning to produce more robust classifiers.

We showed both theoretically and empirically that even under the very general MC learning noise model on the sensitive feature, fairness can still be preserved by scaling the input unfairness tolerance parameter.

PhasePack is a collection of sub-routines for solving classical phase retrieval problems. PhasePack contains implementations of both classical and contemporary phase retrieval routines.

A Tutorial on Fairness in Machine Learning

Reviewer: Neurips2020-2024, ICLR2021-2025, ICML2021-2024, ICRA2023-2024, TOSEM

Lead Teaching Assistant: COMS 4115 Programming Language & Translators(23 Fall)

Teaching Assistant: COMS 4115 Programming Language & Translators(23 Spring), COMS 4771 Machine Learning(18 Summer, 19 Spring), ELEN 4903 Machine Learning(Edx)(18 Spring)