Conference on Cognitive Computational Neuroscience

The question about the nature of representations (that is, about which kinds of entities or processes can be said to be about or stand for other entities or processes) is a central theoretical problem in cognitive science. I present a framework aimed at systematizing the debate around this question. I claim, first, that representations are primarily signals tailored for information transmission, for the sake of information transmission---as opposed to entities that merely happen to carry information. Second, I argue that a way to establish empirically that some entity is a representation, therefore, is to show that it is adapted for information transmission---that it aims at optimally trading-off three key information-transmission budgets: 1) rate (or transmission and storage costs); 2) distortion (or faithfulness of the transmitted information); and 3) computational complexity. Paradigmatic examples of representational phenomena typically sit close to the Pareto frontier of a Rate-Distortion-Complexity multiobjective optimization: the representational manifold of the title.