Keywords: beauty, symmetry, simplicity, processing fluency, logic of justification, logic of pursuit
Abstract: The extrapolative leaps from training to deployment are precisely where ML's best methods for system development succeed, and where most---and the most consequential---failures occur. I argue that the leap is where beauty can and does play a role. Beauty also serves to highlight some important aspects of the leap that are not specific to beauty. Based on this, and drawing on research on beauty from psychology, cognitive science, and philosophy of science, I articulate some fundamental problems and suggest directions for potential solutions.
Category: Meta-research: A meta-research paper on the role of beauty, publishing incentives, role of negative results in ML research