The Inductive Bias of the Architecture

On a simple task like this, i tend to think that almost all the knowledge that you would want out of the problem will be linearily otable from the agent's vision output. And so it's more about finding interesting things by which signals explain more variants in those representations. But i'm also a little bit sceptical of how to interpret those, because they tend to be biased too by the inductive biaces of the architecture, which may not be that rellmet right. In these resentations, we didn't really do any of that, it's totally possible to do.