CV - Jiayou(Rick) Liang / Personal Website

Research experience

4/2025 – Present: Accelerating Adaptive Algorithms At Scale through Reparameterization

Supervisor: Prof. Rebecca Willett (The University of Chicago)
Role: Research Assistant
Implemented scalable approximations of EGOP: built low-rank estimation and periodic reparameterization modules based on auxiliary EGOP and randomized SVD, enabling faster convergence of Adam/Adagrad in large-scale model training without significant computational overhead.
Conducted hyperparameter search and ablation studies on Fashion-MNIST, TinyMNIST, and EMNIST (Letters/Digits), comparing models with/without reparameterization, different optimizers, and learning-rate schedules; reported convergence speed, validation accuracy, memory footprint, and runtime to ensure reproducibility of results.
Related work: Faster Adaptive Optimization via Expected Gradient Outer Product Reparameterization

4/2025 – Present: Deep Assortment Optimization

Supervisor: Prof. Rui Gao (The University of Texas at Austin)
Role: Research Intern
Transformed discrete assortment selection into subset distribution learning, applying continuous relaxations (STGS/SIMPLE/NCPSS/SFESS) and permutation-invariant autoregressive modeling with RNN, enabling differentiable optimization under k-subset constraints.

4/2025 – Present: Foundation Model Fine-tuning for Breast Cancer Biomarker Prediction

Supervisor: Prof. Utku Pamuksuz (The University of Chicago)
Role: Data Scientist
Adapting Direct Preference Optimization from natural language processing to computational pathology through self-supervised preference construction, enabling parameter-efficient fine-tuning of the TITAN foundation model for improved breast cancer biomarker prediction from whole slide images.

Hackthon Project

Host: St. Jude Children’s Research Hospital
We built an LLM agentic chat interface that lets researchers explore the Childhood Cancer Survivor Study (CCSS) via natural-language questions, automatically translating them to SQL and returning instant summary stats and visualizations across 15k+ de-identified variables and two cohorts (survivors, siblings). Deployed a secure PostgreSQL backend with controlled access, and wired it to LangChain agents, and Azure Cognitive Search to accelerate pre-proposal exploration and variable discovery.

CV