Supplementary Material: Unsupervised learning models of primary cortical receptive fields and receptive field plasticity

Independent component analysis (ICA) The ICA algorithm has been applied successfully to modeling V1 simple cell receptive fields [1, 2]. It is closely related to sparse coding methods, and can be cast in terms of a simple generative model [3]: We suppose that our data x ∈ R is an unknown linear mixture of independent, non-Gaussian sources, i.e. x = As where A ∈ Rn×n is unknown. During learning, ICA adapts the weights to unmix these sources. At the end of learning the recovered weights are the pseudoinverse of the mixing matrix A, and the model neuron activities are the values of the independent sources s. We use the standard “FastICA” Matlab implementation [4] which we modified to run on GPU hardware, and we use the log-cosh nonlinearity. The basic ICA method itself has no parameters, and the only parameter we consider here is the number of principal components of the data kept before running the algorithm.