Learning with Mental Imagery


Nov 03, 2017 (modified: Nov 03, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: In this paper, we propose deep convolutional generative adversarial networks (DCGAN) that learn to produce a 'mental image' of the input image as internal representation of a certain category of input data distribution. This mental image is what the DCGAN 'imagines' that the input image might look like under ideal conditions. The mental image contains a version of the input that is iconic, without any peculiarities that do not contribute to the ideal representation of the input data distribution within a category. A DCGAN learns this association by training an encoder to capture salient features from the original image and a decoder to convert salient features into its associated mental image representation. Our new approach, which we refer to as a Mental Image DCGAN (MIDCGAN), learns features that are useful for recognizing entire classes of objects, and that this in turn has the benefit of helping single and zero shot recognition. We demonstrate our approach on object instance recognition and handwritten digit recognition tasks.
  • TL;DR: Object instance recognition with adversarial autoencoders was performed with a novel 'mental image' target that is canonical representation of the input image.
  • Keywords: Deep Learning, Adversarial Networks, Object Instance Recognition, Cognitive AI