A Feature Selection Algorithm for the Regularization of Neuron Models

This paper presents a novel training method for estimating the parameters of retina models, such as integrate-and-fire (IF) or Poisson based. The presented models are constructed using a set of linear and nonlinear filters, which are described by basis functions and Taylor polynomials, respectively. This approach allows for the identification of a set of features that can be used for reproducing retina responses. By using the Bayesian-Laplace feature selection algorithm herein proposed, an efficient model with a reduced set of parameters is achieved. Experimental results show that the proposed algorithm is able to remove unimportant features while still accurately reproducing retina responses. These results also show that the IF model is able to mimic the retina visual processing system using less parameters than the Poisson-based model.