About Me
I'm a third-year Ph.D. student at Mila and the University of Montreal, advised by Professor Aaron Courville.
My life goal is to understand and build general intelligence. Therefore, my research interest evolves as my belief about general intelligence changes.
I'm currently interested in long-context sequence models (particularly linear complexity models) and their potential application in RL.
Previously I've worked on learning object-centric representations using structured generative models with Professor Sungjin Ahn at Rutgers University. I also had experience with computer vision when I started my research career in Professor Xiaowei Zhou's group at Zhejiang University.
About Mind/Intelligence
What I believe:
Eventually, the agent needs online data. In other words, they need to interact with the world. Static datasets are very useful but they have limitations.
The agent should have a memory system. In terms of architecture, this is the most important part.
It is okay to have frozen parameters initially, just like how human DNA sequences are fixed (after birth). The agent can learn to change its own parameters (or even architectures) at some point. This process need not be hardcoded by us. It will emerge.
Thoughts:
About RL: I believe the true value of RL is that it can provide the right ultimate objective to optimize. However, recent advances in deep learning (i.e., LLMs) suggest the best way to optimize this objective might be through currently optimizing some other auxiliary objective, either for obtaining a reasonable initial policy or for getting sufficient learning signal. The main considerations are (1) the amount of data available, (2) the amount of useful learning signal, and (3) how easy it is to optimize the objective.
Language: we think with language (and images) instead of some hidden, internal representations. At least that's what we perceive. This is interesting because this is very inefficient.
