Connecting natural and artificial neural networks with functional brain imaging
CLJ 123 Washington Street ROOM 572
Abstract: Functional neuroimaging and neural network modeling were both introduced to cognitive science in the 1980s, and have both produced influential research. Yet surprisingly, the programs have advanced with little mutual influence. I will describe two different approaches to more directly connecting cognitive neural network models with functional brain imaging in the search for the neural bases of cognition. In the first, direct estimates of gross neural connectivity in real brains are used to shape the architecture of an artificial neural network (ANN).
Simulations with the resulting model are then used to understand patterns of healthy and impaired behaviors, as well as patterns of functional activation revealed by standard univariate contrast methods.
We have applied this approach successfully to develop network models of single-word processing and of semantic cognition that provide a unified account of key phenomena in these domains. Yet this approach also ignores an important contribution of neural network modeling, the possibility that neural representations can be instantiated as highly distributed patterns of activation that are strongly shaped by learning.
In the second approach, I consider how statistical analysis of brain imaging data might best proceed if real neural networks have the properties predicted by neural network models. In analyses of synthetic data generated by such models, we have shown that common univariate and even popular multivariate approaches adopt implicit assumptions that prevent them from discovering essential representational structure. I will describe some new sparsity-based optimization methods that, because they begin by assuming that representations might be distributed, are better able to find such representations where they exist in the data.
When applied to real fMRI data, these methods offer a quite different picture of the nature of neuro-cognitive representations than those yielded by standard approaches. Together the work suggests that artificial neural network models can provide a useful conceptual bridge for connecting cognition and neuroscience, and that the joint application of these methods may lead to new insights about the neural bases of cognition not achievable by other means.