InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Citation

• Authors: Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel
• Year: 2016
• Venue: NeurIPS 2016
• arXiv: arXiv:1606.03657
• URL: https://papers.nips.cc/paper/6399-infogan-interpretable-representation-learning-by-information-maximizing-generative-adversarial-nets

One-Line Takeaway

InfoGAN is a non-VAE but highly relevant disentanglement baseline: it makes selected latent variables semantically active by maximizing a tractable lower bound on the mutual information between those codes and generated samples, rather than relying on an ELBO or KL bottleneck.

Technical Core

• Decompose the generator input into incompressible noise and structured codes: $$z \sim P_{\mathrm{noise}}(z), \qquad c = (c_1, \ldots, c_L) \sim \prod_{i=1}^{L} P(c_i), \qquad x = G(z, c).$$
• Start from the standard GAN minimax objective $$V(D, G) = \mathbb{E}_{x \sim P_{\mathrm{data}}}[\log D(x)] + \mathbb{E}_{z \sim P_{\mathrm{noise}},\, c \sim P(c)}[\log(1 - D(G(z, c)))].$$
• The core failure mode is that a latent-augmented generator can still ignore $c$, so $P_G(x \mid c) = P_G(x)$ and the codes carry no semantic content.
• InfoGAN adds an information regularizer $$\min_G \max_D V_I(D, G) = V(D, G) - \lambda I(c; G(z, c)),$$ so that generated samples should preserve recoverable information about the chosen codes.
• Because $I(c; G(z, c))$ is intractable, the paper introduces an auxiliary recognition model $Q(c \mid x)$ and optimizes the lower bound $$L_I(G, Q) = \mathbb{E}_{c \sim P(c),\, x \sim G(z, c)}[\log Q(c \mid x)] + H(c) \le I(c; G(z, c)).$$
• The final training game is $$\min_{G, Q} \max_D V_{\mathrm{InfoGAN}}(D, G, Q) = V(D, G) - \lambda L_I(G, Q).$$

Method

• Feed both $z$ and structured code $c$ into the generator, with categorical codes for discontinuous factors and continuous codes for smooth ones.
• Parameterize $Q(c \mid x)$ as an auxiliary recognition head that usually shares most convolutional layers with the discriminator, so the extra computation is small.
• Optimize the GAN objective and the variational information lower bound jointly.
• Use DCGAN-style architectures and Adam; the paper reports that $\lambda = 1$ is usually enough for discrete codes and smaller values are often used once continuous codes are included.

Empirical Evaluation

• On MNIST, one categorical code and two continuous codes recover digit identity, rotation, and stroke width.
• On rendered 3D faces and chairs, the learned codes recover pose, lighting, and width-like geometric factors without factor labels.
• On SVHN, the learned factors separate the central digit from contextual or background digits and also capture lighting variation.
• On CelebA, the model discovers codes aligned with pose, eyeglasses, hair style, and emotion.
• The paper also reports that the MNIST categorical code can be matched to digit labels and used as an unsupervised classifier with about 5% error.

Assumptions And Limits

InfoGAN gives a strong empirical recipe for recoverable latent factors, but it does not identify which factors should be discovered or guarantee stable semantics across architectures and code choices. The method is also mostly validated through qualitative traversals, image-domain inductive biases, and carefully chosen latent families, so it should be read as an optimization and inductive-bias story rather than an identifiability result.

Relevance To This Project

This paper is useful for the repo because it isolates a very clean principle: if we want a latent variable to matter, we can reward output-side recoverability of latent information directly rather than only tuning a KL bottleneck. That is close in spirit to the project's inter-latent divergence idea, where different latent values are pushed to leave measurably different traces in the output distribution.

It is also a helpful contrast point for the repo's VAE-centered loss discussions. Beta-VAE and related ELBO variants try to obtain better latent structure by reshaping a bottleneck; InfoGAN instead adds an explicit information objective outside the ELBO. For conditional sequence factorization, the project needs more than recoverable latents because $z$ should represent stable prompt-conditioned strategies, but InfoGAN is a useful precursor for the narrower claim that latent semantics often need explicit pressure rather than hope.

Comparison Notes

beta-VAE and Burgess et al. (2018) are the closest VAE-side analogues, but they encourage disentanglement by tightening or scheduling latent capacity, whereas InfoGAN rewards code recoverability directly. InfoVAE later decomposes the VAE KL into mutual-information and marginal-matching roles; from that angle, InfoGAN is a complementary reminder that latent information can be made explicit instead of merely penalized. In the current repo, inter-latent divergence plays the most similar functional role because it tries to make different latent values visibly different in the output space.

Related Notes

• beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework
• Understanding disentangling in beta-VAE
• Inter-Latent Divergence via Token-Weighted Jensen-Shannon Divergence
• InfoVAE Connections to the Current Loss Structure
• Context