InfoGAN: Interpretable Representation Learning by information maximizing generative adversarial nets
X.Chen, Y.Duan,R.Houthooft,J.Schulman,P.Abbeel, arXiv June 2016.
[code]
In this paper a small modification is done to GAN objective which encourages to learn interpretable and meaningful representations in a unsupervised manner. This is done by maximizing the mutual information between a fixed small subset of GANs noise variables and observation.
Instead of providing one unstructured noise vector to the generator, the input noise vector is decomposed into two parts: incompressible noise and latent code. It is to be noted that the information in the latent code should not be lost in the generation process. The mutual information can be maximized by obtaining lower bound technique which can be easily approximated with monte carlo simulation.
It is implemented by sharing all convolutional layers for D and G with one fully connected layer to output parameters for conditional distribution. Up-convolutional architecture is used for generator networks and leaky RELu with leaky rate 0.1 is used in discriminator hidden layers and normal RELu for generator networks. InfoGAN introduces an extra hyper parameter which is easy to tune.
In this paper a small modification is done to GAN objective which encourages to learn interpretable and meaningful representations in a unsupervised manner. This is done by maximizing the mutual information between a fixed small subset of GANs noise variables and observation.
Instead of providing one unstructured noise vector to the generator, the input noise vector is decomposed into two parts: incompressible noise and latent code. It is to be noted that the information in the latent code should not be lost in the generation process. The mutual information can be maximized by obtaining lower bound technique which can be easily approximated with monte carlo simulation.
It is implemented by sharing all convolutional layers for D and G with one fully connected layer to output parameters for conditional distribution. Up-convolutional architecture is used for generator networks and leaky RELu with leaky rate 0.1 is used in discriminator hidden layers and normal RELu for generator networks. InfoGAN introduces an extra hyper parameter which is easy to tune.