Jin Hwa Lee
Hi, I’m Jin (she/her)!
I am a PhD student at University College London supervised by Andrew Saxe.
At the moment, I am mainly intrigued by compositionality in learning and inference. Compositionality is very general capability that humans have in language acquisition, skill learning and generalization and seems like we might live in the time that artificial intelligent systems exhibit similar ability very soon (or already).
I am interested in:
How do biololgical and artificial networks discover useful motifs or concepts and flexibly reuse them to solve complex tasks?
How do we evaluate machines exhibiting compositional behavior?
How does compositionality emerge during learning?
To answer these questions, I use both theoretical and empirical approach to study.
Before starting my PhD journey, I obtained BSc. Physics at KAIST and I studied Neurosciece for my master, supervised by Mackenzie Mathis and developed multi-modal representation learning method CEBRA for my thesis.
See Research section for more on previous/ongoing projects.
news
| Oct 8, 2024 | We developed a theory about the exact conditions to learn modular representation in neural networks; Don’t cut corners of source distribution! 📐 |
|---|---|
| Oct 2, 2024 | Our project on analyzing phase transition in representational geometry in LLMs during the pre-training phase is out here. |
| Jul 20, 2024 | I’m giving a tutorial talk in Conference on Lifelong Learning Agents! in August, Pisa. I’ll be talking about theoretical approaches in continual learning. |
| Jul 1, 2024 | My first PhD project Why do Animals Need Shaping? A Theory of Task Composition and Curriculum Learning is accepted to ICML 2024! |
| Sep 30, 2023 | I joined Saxe Lab and started a PhD journey 🚀🧠 :)) |