A General Framework for Quantitatively Assessing Neurocomputational Models with FunctionalNeuroimaging Data

Science Center A254

To better understand neural networks, researchers have created computational models of neural networks. DIVA is one of several models which aims to better understand the neural mechanisms involved in speech production. Currently, evaluations of this model has mostly been qualitative, as in researchers look at the models predicted areas of activation and behavioral outputs then subjectively compared them to what would be expected. The problem arises when comparing different models with competing theories as there is no objective qualitative comparison between the two. Functional neuroimaging could potentially be used as a common ground for directly and quantitatively comparing these models. However, there is currently a lack of a framework that will express model neural activity in a way that is directly comparable to empirical data. In this study we begin to develop this framework. The general framework is to generate a Gaussian distribution centered at the locations of each node that when summed together, represent generated full brain activity, will closely approximate the empirical data. Each node of the computational model generates a load value that represents its relative activity during a task at a particular point in time. We use the computational load as a 'switch' to augment the height of each gaussian depending on its activation during that particular task. The center of each gaussian is also variable as the locations of the neural nodes do vary from person to person, it is necessary to move these so that we find the true center for a particular person/averaged data set. The goal then is to optimize a single set of alpha, beta, and location values for three tasks each with differing computational load values. This leaves us with a generated brain activity can be directly compared to empirical data. Future goals will be to further improve the algorithm by testing out differing numbers of nodes and then comparing the DIVA model to other models.