Representation Learning
Studying what makes a visual representation universal, useful, and "good".
News
Research Interests
Vision
Multimodal
Embodied
AI Efficiency
Hardware-software co-design across speculative decoding, quantization, and pruning.
Quantization
Speculative Decoding
Pruning
Agents
Building agent tooling and constructing benchmarks for evaluating agent capabilities.
Benchmark
Tool
Application
Selected Projects
View allResNet Sparse Distillation
The increasing depth and width of neural networks improve accuracy but also raise hardware requirements and slow down inference. This project proposes a distillation loss function that enables immediate weight and activation pruning on the student model after distillation.
JEPA Architecture Probing Model
This project designed a JEPA architecture model with CNN and MLPs as encoder and predictor. The main task is to predict the trajectory of an object in an environment with wall and door.
Blogs
View allFrom SFT to RL: Reward and Policy Gradient
Where RLHF's reward signal comes from, and how policy gradient turns a sequence-level score into token-level updates.
From SFT to RL: The Two Degrees of Freedom
Why SFT is reference-distribution fitting, and how changing token weights and sampling turns it into RL.