ABSTRACT
The description of a neural network’s behavior in terms of its attempts to achieve specific computational goals is necessary to obtain a complete understanding of the network’s behavior. A practical procedure for constructing the computational goal for a given neural network architecture is described. To illustrate the approach, specific computational goals for a variety of neural networks are then formulated as the solutions to a set of nonlinear statistical optimization problems. Some reasons for using a statistical formulation are that (a) rational decision makers should use statistical inference, and (b) a statistical formulation permits an objective evaluation of a neural network’s computational goals using goodness-of-fit tests. Finally, some commonly asked questions about this approach are noted and answered.
