A Rich Source of Labels for Deep Network Models of the Primate Dorsal Visual Stream

Abstract

Deep convolutional neural networks (CNNs) have structures that are loosely related to that of the primate visual cortex. Surprisingly, when these networks are trained for object classification, the activity of their early, intermediate, and later layers becomes closely related to activity patterns in corresponding parts of the primate ventral visual stream. The activity statistics are far from identical, but perhaps remaining differences can be minimized in order to produce artificial networks with highly brain-like activity and performance, which would provide a rich source of insight into primate vision. One way to align CNN activity more closely with neural activity is to add cost functions that directly drive deep layers to approximate neural recordings. However, suitably large datasets are particularly difficult to obtain for deep structures, such as the primate middle temporal area (MT). To work around this barrier, we have developed a rich empirical model of activity in MT. The model is pixel-computable, so it can provide an arbitrarily large (but approximate) set of labels to better guide learning in the corresponding layers of deep networks. Our model approximates a number of MT phenomena more closely than previous models. Furthermore, our model approximates population statistics in detail through fourteen parameter distributions that we estimated from the electrophysiology literature. In general, deep networks with internal representations that closely approximate those of the brain may help to clarify the mechanisms that produce these representations, and the roles of various properties of these representations in performance of vision tasks. Although our empirical model inevitably differs from real neural activity, it allows tuning properties to be modulated independently, which may allow very detailed exploration of the origins and functional roles of these properties.