Skip to content
The methodOpen data
Back to the collection

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Xi Chen, Yan Duan, Rein Houthooft

Show all 6 authors

John Schulman, Ilya Sutskever, Pieter Abbeel

2016 · arXiv

Algorithms
Read the paper
HN score
0
Mentions
6
Readers
5
Months
6
What’s in a score?

From the abstract

This paper describes InfoGAN, an information-theoretic extension to the Generative Adversarial Network that is able to learn disentangled representations in a completely unsupervised manner. InfoGAN is a generative adversarial network that also maximizes the mutual information between a small subset of the latent variables and the observation. We derive a lower bound to the mutual information objective that can be optimized efficiently, and show that our training procedure can be interpreted as a variation of the Wake-Sleep algorithm. Specifically, InfoGAN successfully disentangles writing styles from digit shapes on the MNIST dataset, pose from lighting of 3D rendered images, and background digits from the central digit on the SVHN dataset. It also discovers visual concepts that include hair styles, presence/absence of eyeglasses, and emotions on the CelebA face dataset. Experiments show that InfoGAN learns interpretable representations that are competitive with representations learned by existing fully supervised methods.

Metadata: arXiv. Abstracts retain their original authors’ rights.

In the conversation

Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels

6 selected excerpts · 6 total mentions

kdoherty
Neutral mention · rule
I'm assuming by iterate you mean "generate outputs of a GAN which preserve some order that you can iterate through." The answer is yes, but not in the way that one might think. Despite the fact that we seed the GAN with input noise, there is no guarantee that the GAN makes use of this at all. This is a theme with GANs: we often want to imbue them with prior knowledge that we think is important, but is easily ignored by the GAN. In this case, we want to generate samples from p(x|z), where x is in the space of our data (often images, in this case passwords), but provided it gets good results according the the loss function, your GAN may learn p(x|z)=p(x). This is fine if…
Read the full comment
you don't care about enforcing some relationship between input and generated samples, but here we do. One solution is to use InfoGAN (https://arxiv.org/abs/1606.03657), which adds a term to the loss function that the mutual information between a latent code and the generator output must be high. Your latent code might be drawn from a uniform distribution on [-1,1], and the generator output will be conditioned on this code. This being continuous, it's questionable what "iterate" might mean. On a computer, maybe you iterate through every possible float (as someone mentioned), but if you want to generate N different samples, you could also discretize this distribution to N values on the given interval, each with probability 1/N and sample from this PMF.
Original comment PassGAN: A Deep Learning Approach for Password Guessing
fiter
Neutral mention · rule
The paper I was thinking of is called: "InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets"[0]. I do not have experience training and investigating neural nets, but from what I read in that paper, there's no reason to presume you'll find neurons that represent a feature you're interested in. In the paper they alter the reward function to get neurons that correspond to the features they are interested in. [0] https://arxiv.org/pdf/1606.03657v1.pdf
Original comment Unsupervised sentiment neuron
nl
Neutral mention · rule
GANs (and InfoGAN[0] in particular) has shown there are still pretty huge breakthoughs out there. Some other things you may be interested in: Learning to learn by gradient descent by gradient descent[1] Neural Architecture Search with Reinforcement Learning[2] [0] https://arxiv.org/abs/1606.03657, https://arxiv.org/pdf/1701.00160.pdf [1] https://arxiv.org/abs/1606.04474 [2] https://openreview.net/pdf?id=r1Ue8Hcxg
Original comment Darpa Goes “Meta” with Machine Learning for Machine Learning (2016)
andreyk
Neutral mention · rule
Brief summary: a nice intro about what generative models are and the current popular approaches/papers, followed by descriptions of recent work by OpenAI in the space. Quick links to papers mentioned: Improving GANs https://arxiv.org/abs/1606.03498 Improving VAEs http://arxiv.org/abs/1606.04934 InfoGAN https://arxiv.org/abs/1606.03657 Curiosity-driven Exploration in Deep Reinforcement Learning via Bayesian Neural Networks http://arxiv.org/abs/1605.09674 Generative Adversarial Imitation Learning http://arxiv.org/abs/1606.03476 I think the last one seems very exciting, I expect Imitation Learning would be a great approach for many robotics tasks.
Original comment Generative Models
nl
Neutral mention · rule
They have plenty of results. All their work on GANs[1][2] has been great, and the OpenAI gym has proven to be pretty popular for reinforcement learning research already. I'm not sure how "hype" is measured - perhaps people thought they were going to invent some kind of AI entity we would have to persuade not to shoot us all by now or something. I know that all the papers coming out of OpenAI get hyped just as much as anything from DeepMind, Google Brain or FAIR, so there seems to be decent amount of hype still? But then usually that hype is justified, in which case maybe it is better characterised as "excitement"? [1] http://arxiv.org/pdf/1606.03657.pdf [2] http://arxiv.org/pdf/1606.03498.pdf
Original comment OpenAI Team Update

Keep following the thread.

More in Algorithms

More from the same topics.

2000arXiv

Dancing links

Donald E. Knuth

The author presents two tricks to accelerate depth-first search algorithms for a class of combinatorial puzzle problems, such as tiling a tray by a fixed set of polyominoes.

AlgorithmsProgramming languages
10 mentions10 months
2HN score