Group behavior models for learning in neural networks

A description is given of work on supervised learning algorithms for layered, feedforward neural networks, motivated by group behavior models developed to describe coordinated motion of flocks of birds and schools of fish, for example. The approach involves introducing additional hidden nodes while constraining the associated additional degrees of freedom to avoid excessive computation during the transient phase of learning and to avoid over parameterization asymptotically. If schools are rather widely dispersed. at least on the output side, it is expected that the school learning procedure will offer broader search capabilities because of the extra degrees of freedom available in early stages of the learning process. In the simplest selection strategy, the choice of units within schools is simply a cyclic one. Some caution should be exercised if this is used in conjunction with cyclic presentations of a finite set of training examples