Orientation selective cells emerge in a sparsely coding Boltzmann machine

Investigates a sparse coded Boltzmann machine as a model for the formation of orientation selective receptive fields in primary visual cortex. The model consists of two layers of neurons which are recurrently connected and which represent the lateral geniculate nucleus and primary visual cortex. Neurons have ternary activity values +1, -1, and 0, where the 0-state is degenerate being assumed with higher prior probability. The probability for a (stochastic) activation vector on the net obeys the Boltzmann distribution and maximum-likelihood leads to the standard Boltzmann learning rule. The authors apply a mean-field version of this model to natural image processing and find that neurons develop localized and oriented receptive fields.