Architectural Conditioning for Disentanglement of Object Identity and Posture Information

Kazutoshi Sagi, Takahiro Toizumi, Yuzo Senda

Feb 11, 2018 (modified: Feb 11, 2018) ICLR 2018 Workshop Submission readers: everyone
  • Abstract: This paper proposes an architectural conditioning approach for disentanglement of object identity and posture information. Challenging in deep learning is to disentangle a specific condition from learned representations. The proposed architectural conditioning employs a rigid matrix operation as a layer in an autoencoder to achieve disentangling of a specific condition. This paper demonstrates how the proposed conditioning learns rotation-invariant representations. Using the architectural conditioning, rolling of latent vectors corresponds to rotation of an object in an image. Thus the object posture information is positionally represented in a latent vector. The experimental results on MNIST and 3D chair model images show that this conditioning enables networks to learn rotational bases as their weights. An arbitrary view can be inferred using different views.
  • TL;DR: We propose an architectural conditioning approach to disentangle a specific condition such as image and object rotations in learned representations.
  • Keywords: disentangled representations, condition-limited image, conditional autoencoder