A Method of Automatic Training Sequence Generation for Recursive Neural Networks in the Area of Sensing Directional Motion

In the modeling of vision systems of biological organisms, one of the important features is the ability to sense motion (BorgGraham et al. 1992, Huntsberger 1995, Klauber 1997, Missler and Kamangar 1995, Rosenberg and Ariel 1991). Motion is sensed by animals through neurons that receive input over some area of the field of view (Newman et al. 1982 and Rosenberg and Ariel 1991). For such a neuron to function properly implies the ability to "remember" how things were in the past and combine that information with how things are in the present. In attempting to come up with a computationally efficient model of a motion neuron a Jordan-Elman network has been utilized. The Jordan-Elman network allows for a one time step remembrance of the state space, making it suitable for motion sensing (Jordan 1986 and Elman 1990). To train the network an automatic method of building a training sequence was developed based off earlier work done mostly through hand coding to allow for easy construction of motion sensing networks of differing features. A network with a 3x3 input field and a network with a 5x5 input field were trained using the automatic method of training sequence generation. These two networks were tested against a network with a 3x3 input field trained using the training set constructed partially by hand. Results favored the latter network, but the former networks showed future promise. It is hoped with the right modifications to the creation of their training sequence they will become better than their hand built ancestor.