Feature Learning in Deep Learning
This project studies feature learning in deep neural networks: how useful internal features emerge during training, how they become organized, and how they support generalization.
Technical Research
Technical research on how artificial systems form, organize, and use representations. This includes work on sentence embeddings, language models, canonical task structure, grokking, and the geometry of learned features.
Semantic Geometry of Sentence Embeddings
This paper studies the internal geometry of sentence embeddings, showing how latent semantic features can be identified, composed, and used for more interpretable embedding-based systems.
The Canonical Representation of a Task
This paper proposes that generalization occurs when learned representations align with a canonical representation determined by the task itself, and tests this theory on grokking in modular arithmetic.
The Representation of Space in LLMs
This project studies whether language models build internal representations of space, and how spatial relations might be encoded, composed, and used during reasoning.
Canonical Representations of Language Representations
This project asks whether language representations admit canonical structures: stable representational forms determined by linguistic tasks, semantic relations, or the geometry of meaning itself.