5. Introduction
Related work - DCGAN
Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv preprint arXiv:1511.06434 (2015).
6. Methods
Mutual Information for Inducing Latent Codes
• Input noise vector = z +
c
• z: source of incompressible noise
c: latent code (will target the salient structured sementic features
)
• In standard GAN, generator is free to ignore c by
fi
nding solution satisfying
7. Methods
Mutual Information for Inducing Latent Codes
• Mutual Information
The "amount of information" learned from knowledge of random variable Y
about the other random variable X.
• Cost function
10. Implementation
• Q and D share all convolutional layer
s
• InfoGAN only adds a negligible computation cost to GA
N
• LI(G, Q) always converges faster than normal GAN objectiv
e
• InfoGAN essentially comes for free with GAN
11. Goal: To investigate if mutual information can be maximized ef
fi
ciently
Dataset: MNIS
T
Latent code: uniform categorical distribution
Experiments
Mutual Information Maximization
12. Goal: To evaluate if InfoGAN can learn disentangled and interpretable representations
Dataset : MNIS
T
Latent code:
Experiments
Disentangled Representation
13. Dataset : 3D Face
s
Latent code:
Experiments
Disentangled Representation
14. Dataset : 3D Chair
s
Latent code:
Experiments
Disentangled Representation
16. Conclusion
• InfoGAN is completely unsupervised and learns interpretable and disentangled
representations on challenging dataset
.
• InfoGAN adds only negligible computation cost on top of GAN and is easy to
train
.
• The core idea of using mutual information to induce representation can be
applied to other methods like VAE
