Deep learning methods have been successfully applied to convert high-dimensional to low-dimensional codes . These low-dimensional codes are essentially higher- level features that can provide good discriminability for classification tasks. In this study we train a deep network for feature extraction and classification on the 10,000 dimensional ARCENE dataset. The unsupervised pre-training for feature extraction is done by a stacked autoencoder (SAE), and the subsequent supervised logistic regression by a softmax classifier. We then compare the balanced error rate (BER) performance measure of the deep network with that of a feedforward neural network. The report also highlights why a Bayesian neural network (BNN) was considered but not selected for the study.
Full report: paper.pdf