PhD Seminar • Machine Learning | Applied Category Theory • Reconciling Auto-Differentiation with Traditional Backprop, and Building Neural Networks

Please note: This PhD seminar will take place in DC 3317.

Nolan Peter Shaw, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Jeff Orchard

In this talk, I present a new categorical object, pre-lenses, and show how they unify the two perspectives of performing backpropagation (all forward then all back, vs. forward and back together as is done with autodifferentiation). I will also discuss how these pre-lenses allow us to reinterpret parametric lenses such that they can be described as profunctors. This allows neural network construction using nothing more than vanilla function composition.