Doina Precup

We show that deep neural networks across architectures and training conditions all instantiate the same abstract algorithm for modular addition.

We provide a formal definition of plasticity under minimal assumptions, and highlight a fundamental connection to empowerment