Sparse hierarchical models of vision

Abstract/Summary

In recent years, deep convolutional neural networks (CNNs), which are based on the hierarchical structure of the visual cortex has found remarkable accuracy rates in object classification. One of its drawbacks is the requirement of a large collection of labelled data for training. Therefore, unsupervised hierarchical networks are more suited for a more biologically plausible model. For a classification accuracy to be closer towards CNNs, the invariance and selectivity of the extracted features need to be improved. One of the standard methods for learning invariant features is to apply non-linearity functions to the data which has been implemented in both CNNs and HMAX models. Based on these principles, an extended form of the HMAX model is proposed which applies two different types of non-linear pooling operations. The extension is designed with the help of sparsity based algorithms such as Independent Subspace Analysis (ISA) and Topographic Independent Component Analysis(TICA). Aside from an improved classification accuracy compared to previous unsupervised hierarchical models, it also reduces the data dimensions within the layers.