ABSTRACT
In addition to addressing specific patterns of behavioral and neural data, neural network models are valuable for their ability to establish general principles of functional neural organization. In particular, computational models can explain how differences in the structuring and parameters of neural networks can lead to qualitatively different, often mutually incompatible, capabilities. To the extent that different computational capacities require fundamentally different kinds of neural networks, the brain could have either a compromise or a trade-off between the different network properties, or it could specialize different brain areas for different functions to avoid such a trade-off. Critically, this kind of computational approach to functional neural organization enables one to understand both what is different about the way different neural systems learn, and why, from a functional perspective, they should have these differences in the first place. Thus, the computational approach can go beyond mere description towards understanding the deeper principles underlying the organization of the cognitive system.
