Proceedings of the Annual Meeting of the Cognitive Science Society

Efficient data compression is essential for capacity-limited sys-tems, such as biological memory. We hypothesize that the needfor efficient data compression shapes biological perception andperceptual memory in many of the same ways that it shapesengineered systems. If true, then the tools that engineers useto analyze and design systems, namely rate-distortion theory(RDT), can profitably be used to understand perception andmemory. To date, researchers have used deep neural networksto approximately implement RDT in high-dimensional spaces,but these implementations have been limited to tasks in whichthe sole goal is compression with respect to reconstruction er-ror. Here, we introduce a new deep neural network architecturethat approximately implements RDT in a task-general manner.An important property of our architecture is that it is trained“end-to-end”, operating on raw perceptual input (e.g., pixels)rather than an intermediate level of abstraction, as is the casewith most psychological models. We demonstrate that ourframework can mimick categorical biases in perception andperceptual memory in several ways, and thus generates spe-cific hypotheses that can be tested empirically in future work.