Feature discovery by deep learning for aesthetic analysis of evolved abstract images

We investigated the ability of a Deep Belief Network with logistic nodes, trained unsupervised by Contrastive Divergence, to discover features of evolved abstract art images. Two Restricted Boltzmann Machine models were trained independently on low and high aesthetic class images. The receptive fields (filters) of both models were compared by visual inspection. Roughly 10 % of these filters in the high aesthetic model approximated the form of the high aesthetic training images. The remaining 90 % of filters in the high aesthetic model and all filters in the low aesthetic model appeared noise like. The form of discovered filters was not consistent with the Gabor filter like forms discovered for MNIST training data, possibly revealing an interesting property of the evolved abstract training images. We joined the datasets and trained a Restricted Boltzmann Machine finding that roughly 30 % of the filters approximate the form of the high aesthetic input images. We trained a 10 layer Deep Belief Network on the joint dataset and used the output activities at each layer as training data for traditional classifiers (decision tree and random forest). The highest classification accuracy from learned features (84 %) was achieved at the second hidden layer, indicating that the features discovered by our Deep Learning approach have discriminative power. Above the second hidden layer, classification accuracy decreases.