Insu Jeon
Ask or search…
Comment on page


Disentangled Representation Learning

  • Disentangled representations: a change in a single direction of the latent vector corresponds to changes in a single factor of variation of data while invariant to others.
  • GOAL: Learning an Encoder that can predict the disentangled representation. Learning a Decoder (or Generator) which can synthesize an image.
  • HARD: Achieving the goal without truth generative factors or supervision is hard.

Information Bottleneck (IB) Principle

  • GOAL: Obtaining the optimal representation encoder q_ϕ (z│x) that balances the trade-off between the maximization and minimization of both mutual information terms.
  • I(⋅,⋅) denotes mutual information (MI) between input variable X and target variable Y.
  • The learned representation Z acts as a minimal sufficient statistic of X for predicting Y.

Information Bottleneck GAN

IB-GAN introduces the upper bound of MI and 𝛽 term to InfoGAN’s objective, inspired by IB principle and 𝛽-VAE for the disentangled representation learning.
𝑰^𝑳 (⋅,⋅) and 𝑰^𝑼 (⋅,⋅) denote the lower and upper-bound of MI, respectively (𝜆 ≥𝛽). IB-GAN not only maximizes the shared information between the generator 𝐺 and the representation 𝑧 but also allows control of the maximum amount of information shared by them using 𝛽 analogously to that of 𝛽-VAE and IB theory.


Variational Lower-Bound

  • The lower bound of MI is formulated by introducing the variational reconstructor 𝒒_𝝓 (𝒛|𝒙). Intuitively, the maximization of MI is achieved by reconstructing an input code 𝑧 from a generated sample 𝐺(𝑧)=𝑝_𝜃 (𝑥│𝑧), similar to the approach of InfoGAN.

Variational Upper-Bound

  • Naïve variational upper-bound of generative MI introduces an approximating prior 𝒅(𝒙). However, any improper choice of 𝒅(𝒙) may severely downgrade the quality of the synthesized sample from generator 𝒑_𝜽 (𝒙|𝒛).
  • We developed another formulation of variational upper-bound of MI term based on the Markov property: if any generative process follows 𝑍→𝑅→𝑋, then 𝐼(𝑍,𝑋)≤𝐼(𝑍,𝑅). Hence, we use an additional stochastic model 𝑒_𝜓 (𝑟│𝑧). In other words, we let 𝐺(𝑟(𝑧)).

IB-GAN Architecture (tractable approximation)

  • The IB-Gan introduces the stochastic encoder 𝑒_𝜓 (𝑟│𝑧) before the generator to constrain the MI between the generator and the noise 𝑧.
  • IB-GAN is partly analogous to that of 𝛽-VAE but does not suffer from the shortcoming of 𝛽-VAE generating blur image due to MSE loss and large 𝛽≥1.
  • IB-GAN is an extension of InfoGAN, supplementing an information-constraining term that InfoGAN misses, and shows better performance in disentanglement learning.


Example of generated images from IB-GAN in the latent traversals experiment [1]. (a) IB-GAN captures many attributes on the CelebA [2] and (b) 3D Chairs dataset [3].
  • Comparison between methods with the disentanglement metrics in [1,5]. Our model’s scores are obtained from 32 random seeds, with a peak score of (0.826, 0.74). The baseline scores except InfoGAN are referred to [6].


  • IB-GAN is a novel unsupervised GAN-based model for learning disentangled representation. The IB-GAN's motivation for combining the GAN objective with the Information Bottleneck (IB) theory is straightforward. Still, it provides elegant limiting cases that recover both the standard GAN and the InfoGAN.
  • The IB-GAN not only achieves comparable disentanglement results to existing state-of-the-art VAE-based models but also produces a better quality of samples than standard GAN and InfoGAN.
  • The approach of constructing the variational upper bound of generative MI by introducing an intermediate stochastic representation is a universal methodology. It may advance the design of other generative models based on the generative MI in the future.