0
$\begingroup$

Mode collapse is a known issue in generative adversarial networks (GANs) whereby the generator only learns a subset of the real data distribution. In those cases, it only outputs variations of a small number of images. For example, in the MNIST handwritten digit dataset, the generator might only learn to produce 7s and 9s.

In my experiments, the generator replicates similar square patches of pixels within the same output image, and not necessarily across multiple images. Has this phenomenon been studied in GANs, and is it still called mode collapse?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

The generator of a GAN learns to create images that the discriminator cannot distinguish from real data. Mode collapse just means that the generator, unfortunately, creates only images that are equal to a small subset of the real data. But the generated data is still very similar to some real data.

But the phenomenon you are describing cannot be explained like that: the discriminator checks the generated image for similarity with some real images, it does not check parts of the generated image for similarity with other parts of the same image. Replicating patches within the same image would not result in images that equal some of your real images, so the discriminator would flag them as fake.

So unless you happen to have such images with replicating subregions in your dataset of real images, I am afraid it is probably rather a bug :(

Improve this answer
$\endgroup$
1
  • $\begingroup$ Hi, thank you for your answer! In hindsight, using "modal collapse" anywhere in my question was probably ill-advised because of its association with repeating similar modes of the data, and not so much patterns/qualities. To address your answer, however, I doubt that there is a bug, since I'm really just using a very "vanilla" DCGAN (almost the same as the one in the Torch documentation). I actually suspect that the generator may be underpowered in the first couple layers, since it most noticeably reproduces similar patches in an 8x8 pattern. $\endgroup$
    – Saucy Goat
    Commented Jun 6, 2022 at 8:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.