Neural Population Coding of Stimulus Features

While empirical evidence suggest that the brain can represent and operate on probability distributions, it is not clear how multivariate dependencies can be detected and represented by neural circuits. Based on previous work and the principle of entropy distillation, the paper introduces a massively parallel connectionist machine whose spiking behavior adapts to the statistical distribution of binary inputs. Experimental results confirm that the network is able to accurately capture the joint probability distribution of the inputs and it is able to represent even higher-order features lacking pairwise correlations.