Brain and Cognitive Science Department, Rochester University
A Neural perspective on population coding
Many sensory and motor variables in the cortex are encoded through the activity of large populations of neurons with bell-shaped tuning curves. The optimal way to read out these population codes is to use a maximum likelihood estimator. It is optimal in the sense that it recovers all the Fisher information available in the population activity; it is therefore equivalent an ideal observer.
The question we address is whether cortical circuits can behave like ideal observers. We show that the answer is yes; i.e., cortical circuits have the required connectivity and nonlinearity to behave like a close approximation to maximum likelihood estimators. In other words, each layer in the cortex can be an ideal observer of the activity coming from the preceding layer.
This result has important implications for the theory of attention. Attention is believed to affect the responses of cells by either decreasing the widths of tuning curves or increasing the gain. It is often argued that these changes could account for the behavioral improvements that have been observed under attentive conditions. These conclusions, however, are drawn from models using distinct components for the ideal observer and the cortical circuitry. We will show that they do not necessarily hold when the cortical circuitry and the observer are the same.