Shared visual illusions between humans and artificial neural networks

Alan Stocker

Abstract: Any information processing system should allocate re- sources where it matters: it should process frequent vari- able values with higher accuracy than less frequent ones. While this strategy minimizes average error, it also intro- duces an estimation bias. For example, human subjects perceive local visual orientation with a bias away from the orientations that occur most frequently in the natu- ral world. Here, using an information theoretic measure, we show that pretrained neural networks, like humans, have internal representations that overrepresent frequent variable values at the expense of certainty for less com- mon values. Furthermore, we demonstrate that optimized readouts of local visual orientation from these networks’ internal representations show similar orientation biases and geometric illusions as human subjects. This surpris- ing similarity illustrates that when performing the same perceptual task, similar characteristic illusions and bi- ases emerge for any optimal information processing sys- tem that is resource limited.