Jason Morton

As part of an effort to bring mathematicians and computer scientists working in the area together to explore and strengthen the mathematical foundations of deep learning, I organized organizing a two-day workshop on the topic. The purpose of the workshop was to discuss strategies for improving these foundations, including but not limited to

• formal statements and proofs of the "deep architectures are better" hypothesis,
• mathematical strategies for a priori comparison of architectures for learning tasks,
• questions about deep belief networks that can be translated to problems in algebraic and tropical geometry, such as identifiability and singular learning theory,
• inference functions, encodings, and circuit complexity, and
• how all of the above relates to learning algorithms.

Thanks to the wonderful participants, the workshop led to a lot of interesting new ideas and research directions.