Evaluating network performance: A question of transparency under analysis
Pages 21

The main point that emerges from our discussion so far is that in order to fully evaluate the theoretical implications of a connectionist simulation, the network must be transparent under analysis (McCloskey, 1991). That is, we must be able to determine what factors underlie network performance so that we can understand how the network does what it does. There are several possibilities that arise when interpreting the performance of the FM91 simulation, e.g:

• The network’s performance might be solely determined by parameters of the simulation that are not specified by the theoretical hypothesis under examination. In the present case these include the learning algorithm, the number of units in the network, the weight update procedure, the nonlinear activation function applied to units in the network, the scoring procedure, and the number of concepts on which the network was trained.3