Connectivity and architecture
We studied how a model’s architecture affects the geometry it learns in input space. CNNs and Transformers build in different assumptions about how information is combined. We wanted to understand how these inductive biases show up in the structure of the regions formed by a trained model.
Connectivity and topology
Our main focus was connectivity: whether a region forms a connected set or separates into distinct components. We compared this structure across architectures and also studied other topological properties of the learned regions.