Neural networks as a critical level of description for cognitive neuroscience

With the success of artificial neural network models in machine learning has come a renewed interest in the possibility that neural networks can be used as scientific models for understanding the function of real neural systems. When similar questions initially arose in the 1980’s and ‘90’s, the many discrepancies between artificial and natural neural systems contributed to the widespread view that neural network models were not biologically plausible, and hence of limited utility for understanding real neural systems. The current paper suggests, to the contrary, that such models capture a level of description isomorphic to that adopted by two essential tools in cognitive neuroscience: functional brain imaging and connectivity analysis. Recognizing this concordance allows neural network models to serve as conceptual bridges between hypotheses about neuro-cognitive mechanisms and the structural and neurophysiological measurements that are the raw stuff of cognitive neuroscience. To illustrate these points, the paper reviews four different ways in which neural network models are reshaping our understanding of the neural systems that support human semantic memory.