Geometry of visual representations : each face is a vector

How can we encode a mental representation using a neural vector in a high-dimensional space ? The lecture will take the example of visual recognition : every object, every face we recognize is encoded by the activity of a population of neurons in the inferotemporal cortex. Doris Tsao's work has taken this idea a step further : it shows how each face we recognize is broken down into some fifty main axes of shape and contrast, which can be identified at the neural level. Our knowledge of the neural code of faces is now such that, on the basis of a recording of neural activity, we can reconstruct a composite portrait of the face the animal saw ! The lecture will introduce the concepts of " metamers " (different faces leading to the same neuronal activity, at least in some neurons) and discuss the possibility of " adversarial attacks " (in which one image is slightly modified so that it is confused with another) in machines as well as in the brain.