Guest Lecture: Dr. Thomas Langlois

Revealing the geometry of visual memory encoding using transmission chains

Thomas Langlois (University of California, Berkeley and Princeton University)

Abstract: The visual system encodes information selectively due to limited resources, resulting in systematic distortions and biases. Previous work has revealed systematic biases in spatial memory representations. Spatial memory tasks explicitly require encoding a piece of information (such as the location of a point) with respect to others (visual regions in an image or a scene). In this paper, we show that iterating this simple task greatly amplifies these biases, providing unprecedented insight into the geometry of visual encoding representations. Our results confirm that differences in visual discrimination acuity (change sensitivity) over the same image regions are predictive of the structure of these representations, and that they change as a function of the duration of temporal encoding. We also show that visual encoding representations revealed by this novel experimental paradigm support visual recognition for humans, and that they are similar to biases in convolutional neural networks trained for large-scale object recognition.