My name is Yihong Chen, a researcher affiliated with OATML. I research on AI knowledge acquisition, specifically on how different AI systems can learn to abstract, represent, and use concepts/symbols efficiently. Particularly I care about bridging the gap between structured AI paradigm (e.g. symbolic AI) and unstructured AI paradigm (e.g. large language models) so that we can have a unified understanding of how knowledge formation happens in intelligent units. My thesis articulates the vision.
With this vision in mind, my recent focus is on AI reasoning, memory and continual learning. If you would like to get in touch, you can reach me at yihong-chen AT outlook DOT com. If you are a student at Oxford, please see my professional page here for opportunities to get involved.
π₯ Jul 2025, My PhD thesis is out! Knowledge engines need not just structure, but also destructuring β for plasticity, flow, and adaptability.
π₯ Mar 2025, We have a new tool for interpreting transformers, check out jet expansion.
π₯ Mar 2024, Quanta Magazine covers our research on active forgetting. Check out the article here.
π₯ Dec 2023, I will present our forgetting paper at NeurIPS 2023. Check out the poster here!
π₯ Sep 2023, I presented our latest work on forgetting at IST-Unbabel seminar.
π₯ Jul 2023, I presented our latest work on forgetting in language modeling at ELLIS Unconference 2023. The slides are available here. Feel free to leave your comments.
π₯ Jul 2023, discover the power of forgetting in language modeling! Our latest work, Improving Language Plasticity via Pretraining with Active Forgetting, shows how pretraining a language model with active forgetting can help it quickly learn new languages. You'll be amazed by the model plasticity imbued via pretraining with forgetting. Check it out :)
π₯ Nov 2022, our paper, REFACTOR GNNS: Revisiting Factorisation-based Models from a Message-Passing Perspective, will appear in NeurIPS 2022! If you're interested in understanding why FMs can be some special GNNs and make them usable on new graphs, check it out!
π₯ Jun 2022, if you're looking for a hands-on repo to start experimenting with link prediction, check out our repo ssl-relation-prediction. Simple code, easy to hack π