Fan Feng

I am a postdoctoral researcher at UCSD and the CMU/MBZUAI CLeaR group, working with Kun Zhang and Biwei Huang. Previously, I also work closely with Sara Magliacane at the University of Amsterdam. I completed my PhD at City University of Hong Kong, working with Rosa Chan and closely collaborating with Qi She (Bytedance).

My long-term research goal is to build agents that not only imagine the world but also understand it, act in it, discover goals within it, and continually refine their internal models in an open-ended and self-improving loop. To achieve this, my current research centers on generative learning for causal world models and reinforcement learning agents. Specifically, my research therefore focuses on the following questions:

Can we encode passive or offline observations into meaningful representations that are minimal yet sufficient world models for control and planning, especially with domain or distribution shifts?
How can agents compositionally reuse learned structure in new tasks and environments?
How can agents actively explore in a purposeful, open-ended way that discover and achieve the goals, and also improve the world model?

Beyond research, I am also actively engaged in community building. I co-organize the “I Can’t Believe It’s Not Better” (ICBINB) workshop series, which brings together researchers to discuss the practical challenges, failure modes, and lessons learned in deploying ML systems, to promote a culture of open, constructive discourse around what does not work and how we can collectively build more robust and reliable learning systems:)

Feel free to contact me for research collaborations or other engagements.

Github / Twitter / Google Scholar

ffeng1017 [at] gmail.com

Publications (* marks equal contribution.)

🌟 Selected 📚 Topic: Causal Reinforcement Learning | Structured Representation Learning | Continual Learning for Robotic Vision

Fan Feng

Publications (* marks equal contribution.)

Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making

Provably Learning Task-Relevant World Representation

Learning Interactive World Model for Object-Centric Reinforcement Learning

Online Time Series Forecasting with Theoretical Guarantees

Null Counterfactual Factor Interactions for Goal-Conditioned Reinforcement Learning

Towards Empowerment Gain through Causal Structure Learning in Model-Based Reinforcement Learning

Causal Information Prioritization for Efficient Reinforcement Learning

Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations

Learning Dynamic Attribute-factored World Models for Efficient Multi-object Reinforcement Learning

Factored Adaptation for Non-Stationary Reinforcement Learning

AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning

Community Channel-Net: Efficient Channel-wise Interactions via Community Graph Topology

Power Law in Deep Neural Networks: Sparse Network Generation and Continual Learning with Preferential Attachment

Towards Lifelong Object Recognition: A Dataset and Benchmark

OpenLORIS-Object: A Robotic Vision Dataset and Benchmark for Lifelong Deep Learning

Service

Talks