Evaluation of Autoencoders for Bases to Represent Neuroimaging Data

Ashish Gupta, Murat Seçkin Ayhan, Anthony S Maida

Abstract: Computer vision and machine learning methods have been shown to be more effective than clinicians in analyzing neuroimages since they eliminate human factors like subjectivity and experience. Nevertheless, much of the study involves the extraction of hand–crafted features that are hard to visually interpret. Recently, a sparse autoencoder was shown to be effective in learning feature detectors (bases) by which new representations of neuroimages are obtained. However, no comparison was made among the variants of autoencoders. In this paper, we investigate the advantages of constrained autoencoders: sparse, denoising and contractive, over basic autoencoder. We found that given a large enough basis set, the constrained autoencoders have no significant advantage, in terms of classification performance, over the simple autoencoder. However, the constrained autoencoders significantly reduce the number of bases. As a result, they enable lower dimensional representations.