Neurocomputation of image motion

The Shannon information theory is used to analyze a simple three-layer feedforward neural net. The central assumption is that the transmission channel from the top layer of the network to the middle layer has a short-term memory of the past history of the incoming visual signal. The criterion for the emergence of motion detection is to demand that the output signals from the bottom layer sent signals belonging to statistically independent events in the output signal space. It is shown that the network operates with different modes. One active mode is identified to have the ability to detect the temporal changes of the visual signals and the direction of the image motion. These are shown analytically