The Arms Race of Models: Complexify or Die

Abstract

As a graduate student at the Center for Adaptive Behavior and Cognition (ABC), I learned that simple models can beat more complex models, particularly in out-of-sample accuracy. This finding has been validated in my career since then. But I did observe some situations where complex models were useful. And even more surprisingly, I noticed some situations where complex models were truly necessary. The model was driven to complexity by an arms race. I will illustrate such an arms race with a series of diagrams showing how a classifier can be driven to complexity, focusing on the case of spam.

Talk at “Decision Making Under Radical Uncertainty”, June 11-12, 2021. Berlin, Germany (digital)
Talk recording: https://vimeo.com/568707053