Decontextualized learning for interpretable hierarchical representations of visual patterns

Abstract

Apart from discriminative modeling, the application of deep convolutional neural networks to basic research utilizing natural imaging data faces unique hurdles. Here, we present decontextualized hierarchical representation learning (DHRL), designed specifically to overcome these limitations. DHRL enables the broader use of small datasets, which are typical in most studies. It also captures spatial relationships between features, provides novel tools for investigating latent variables, and achieves state-of-the-art disentanglement scores on small datasets. DHRL is enabled by a novel preprocessing technique inspired by generative model chaining and an improved ladder network architecture and regularization scheme. More than an analytical tool, DHRL enables novel capabilities for virtual experiments performed directly on a latent representation, which may transform the way we perform investigations of natural image features, directly integrating analytical, empirical, and theoretical approaches.