Bayesian Decoding As A Testing and Development Link Between Behavioral and Neurological Models
Formal models of behavior express or explain an expected relationship between stimuli and responses in a controlled (repeatable) experiment. Formal models in the behavioral neurosciences (more often, "computational models") describe, at one or more levels, such as neurotransmitter uptake, single neuron spike trains, or regions of the brain, the neurochemical events that lead to an activity downstream in the network, which is sometimes but not always the observable response. In both fields, the question why a given response is observed on a given trial is answered by reproducing the behavior or activity (statistically) in the model. At least with respect to cross-communication and mutual benefits, there are two well-known problems with this shared paradigm. First, behavioral models are specified at a level of abstraction far beyond the level of neurological models. It is not obvious, therefore, whether there should be any relationship at all between the two constructions, or whether an apparent similarity is meaningful or constitutes 'evidence' to support the two theories. Second, neurological activity, even at the level of single neurons, is exceedingly complex, and this complexity is observable in minute detail. The idea of limiting the number of free parameters so that goodness-of-fit statistics can be compared is self-defeating from the outset. In this paper/proposal, I will briefly explain how a relatively new statistical methodology in the neurosciences, called Bayesian decoding, can connect the two sciences in a rigorous manner. Basically, instead of asking what events give rise to intelligent behavior, the modeler asks what information about the stimulus and the possible responses is contained in the events that lead to the response, that is, what properties of these events can be predicted and to what degree.