Abstract
Serial reproduction reveals that human memory for the 3D views of everyday objects tends to be biased towards canonical views, such as the top, bottom, front, back and side views of the object. We propose a theory of human memory for 3D views based on Shannon’s classical communication model that explains these biases in terms of the efficient compression of visual percepts into discrete categories that are transmitted by a speaker to a listener. We show that incremental memory biases revealed by serial reproduction are predicted by a model that efficiently compresses perspective views into view categories by optimizing the information bottleneck (IB) trade-off between the complexity and accuracy of a lexicon of possible view categories. Our results challenge the frequency hypothesis, which argues that canonical views for an object correspond to the views that are the most commonly encountered for that object. Instead, our proposal suggests that canonical views correspond to the optimal compression of all possible views into a discrete set of views that reflects the optimal trade-off between the complexity and accuracy of the representation.