A Probabilistic Network Model of Population Responses

A central question in computational neuroscience concerns how the response properties of neural populations depend on the activity of neurons both within and outside the population. Various models contain different types and combinations of feedforward and recurrent connections. Each model can be characterized by the particular set of assumptions about what information underlies the population response, and these in turn are reflected in the model’s behavior. We propose that the population response is designed to preserve full information about the relevant dimension in the stimulus, which could be a single unambiguous value, a single ambiguous value, or more than one value. We design an objective based on preserving this information, and use it in training the weights in a model. Our results demonstrate that a combination of feedforward and recurrent connections can generate broadly information-preserving population responses in a network.