A network that uses few active neurones to code visual input predicts the diverse shapes of cortical receptive fields

Paper de Fritz.

This paper refines the method used by Olshausen to generate RFs similar to the ones found in V1. To do so, they use ‘hard sparseness’ (instead of ‘the soft sparseness used by Olshausen):

The agenda in this paper is to compare models of efficient coding that incorporate either hard or soft sparseness in their ability to predict receptive fields recorded in primary visual cortex (Ringach, 2002). We investigate the Sparsenet (Olshausen and Field, 1996) as an example of a model that uses soft sparseness and two different models that enforce hard sparseness. The first is the “sparse-set coding network”, a novel model that explicitly optimises the sparse selection of active neurones to achieve efficient coding. The second model serves as a control; it is a naive combination of Sparsenet with a mechanism for pruning small activity values.

Very interesting and clear paper. The intro is long but very easy to follow and the methods are very well explained. Their conclusion is that a model based on ‘hard sparseness’ is metabolically more effcient (they calculate the number of spikes required by each method) and produces more realistic RFs (they compare they results to the data from Ringach 2002).