Internal representation of a neural network that detects local motion

This paper describes a three-layered neural network that detects local motion from two successive images. The perception of local motion is considered to consist of two stages computationally, that is, the generation of local motion candidates and their interactive selection for true motion detection. After the learning process, the recognition rate for natural images is 99.5% for learned patterns and 89.8% for unknown patterns. The network obtained the algorithm of the two stages automatically. The internal representation for the first stage is implemented as connection weights from input to hidden layers, which matches the function of on-center or off-center cells. The internal representation for the second stage is implemented as connection weights from hidden to output layers which corresponds to lateral inhibition.