Evaluating Sparse Codes on Handwritten Digits

Sparse coding of visual information has been of interest to the neuroscientific community for many decades and it is widely recognised that sparse codes should exhibit a high degree of statistical independence, typically measured by the kurtosis of the response distributions. In this paper we extend work on the hierarchical temporal memory model by studying the suitability of the augmented spatial pooling (ASP) sparse coding algorithm in comparison with independent component analysis (ICA) when applied to the recognition of handwritten digits. We present an extension to the ASP algorithm that forms synaptic receptive fields located closer to their respective columns and show that this produces lower Naive Bayes classification errors than both ICA and the original ASP algorithm. In evaluating kurtosis as a predictor of classification performance, we also show that additional measures of dispersion and mutual information are needed to reliably distinguish between competing approaches.