Bayesian estimates from heterogeneous population codes

There is growing evidence that aspects of perception and behavior can be described as Bayesian inference. Consequently, the ability to perform Bayes-optimal estimation of a stimulus encoded in a neural population can serve as a principle for evaluating the optimality of a neural representation. Here we show that the center-of-mass (COM) decoder can produce estimates of encoded stimulus parameters that are consistent with Bayesian estimates. We predict that a neural system that uses a COM decoder to implement a Bayesian estimator encodes the likelihood function in the shape of the tuning curves and the prior distribution in the preferred stimulus values. Bayesian estimation using the COM decoder is suggested as a principle for explaining the representation of sound source direction in the owl's auditory system.