A neural network model of motion detection for moving plaid stimuli

The perception of moving plaid stimuli is examined using a recurrent neural network trained on the intersection-of-constraints. It is found that the nodes in the network develop representations similar to neurons seen in the animal literature. Although motion signals from 2D-intersections are available, early layers in the network develop component directional-selectivity similar to neurons in VI of the macaque monkey. Nodes higher up in the network show both object directional-selectivity and component directional-selectivity, similar to neurons found in MT of the macaque. These results support the notion of a two stage motion system which relies on component motions rather than feature tracking