Friday, 17th November - 11am Siddharth Narayanaswamy : Seminar

Title:   On the utility of explicit structure for representation learning

Abstract: 

We've seen some fantastic development in the last few years on the ability to generate in the perceptual domains. Driven fundamentally by the pattern-learning abilities of transformers, they are astonishingly adept at generating quality output. However, this has also meant that the implicit/soft compositional nature of transformers makes the entire operation of the model an opaque black box.

In this talk, I'll present some recent and ongoing work that explores the utility of simple but powerful structural inductive biases, which, when employed judiciously, enables a framework that can start learning useful multi-level representations with orders of magnitude fewer parameters than corresponding transformer-based models. We hope that these initial steps and concomitant findings allow exporation of further capabilities that lead to more interpretable and explainable models of perception.