4 ms·

I think it's a bit misleading to claim that AI is "generating" these images, as if from whole cloth. Images of real people have been scored along the adjustabl

• thomasahle 8y ago
  In the recent Deep Mind paper on Big GANs, they discuss this point and compare their generated images with the most similar images in the training set. They certainly makes it seem like the network has learned something real and can generate new images that are not just blends of previous ones.
  • browsercoin 8y ago
    yeah the progress is scary. it will possible at some point to have our AI generate highly addictive binge worthy tv shows real time. imagine if you could watch breaking bad season 9 for example, Walter White Jr. breaks bad
• Yreval 8y ago
  This is an incorrect description of what the algorithm is doing. Two networks--a "generator" and a "discriminator"--play a minimax game where the generator maps random vectors into images, and the discriminator attempts to distinguish between these images and real images of celebrities. When the discriminator is (close to) optimal, image-gradients based on the features it uses to predict are passed to the generator, so that the generator can learn the distribution in feature space that makes up human faces. > Images of real people have been scored along the adjustable metrics, and then as you click the +/- adjustments, the source images are very craftily "blended" to produce a finite spectrum of results. There is no reason to believe the source images are actually embedded in the latent space. The descriptors that you can edit in the gui don't necessarily span an orthogonal basis in that latent space, so some of them are correlated, which is why editing one value can change others. Additionally, there is no a priori reason to believe that the manifold of "human face-like images" of 628x1024 is 512-dimensional, so there are areas of the space that still don't map well to real images. The network's ability to cover this space is limited by the number of unique training images it sees, how long it is trained, and its architecture.
  • greeneggs 8y ago
    > The descriptors that you can edit in the gui don't necessarily span an orthogonal basis in that latent space, so some of them are correlated, which is why editing one value can change others. I think both you and the author of the article are making the same mistake here. (Although at least you use "orthogonal" and "correlated," whereas the author calls nonorthogonal vectors "entangled" for some reason.) If you have a nonlinear function f on a vector space, there's no reason why an orthogonal basis for that space will give a better parameterization than a nonorthogonal basis. Even if you have a linear function, there's no reason why that should make a difference. (For example, take f(x,y) = (x-y,y). Then f(x,0)=(x,0) and f(y,y)=(0,y), so "correlated" input directions (1,0) and (1,1) are mapped to "independent" or orthogonal outputs.) I think it is a bit of a mystery why Gram-Schmidt orthogonalization makes a difference here. Perhaps the author should experiment more with different inner products.
    • Yreval 8y ago
      >I think both you and the author of the article are making the same mistake here. Maybe? >If you have a nonlinear function f on a vector space, there's no reason why an orthogonal basis for that space will give a better parameterization than a nonorthogonal basis. I don't think I made that claim. Here's all I'm saying: To whatever degree the features of interest are linearized in the latent space (and there's really no guarantee that they are), we don't have any guarantee that those linear features are orthogonal to one another, so tuning the latent representation along one feature will also impact others. > (For example, take f(x,y) = (x-y,y). Then f(x,0)=(x,0) and f(y,y)=(0,y), so "correlated" input directions (1,0) and (1,1) are mapped to "independent" or orthogonal outputs.) That's true, but remember that the nonlinear mapping is from our latent space (spanned by uniformly random 512-element input vectors) to pixel space. We really don't care about linear algebra in pixel space. I have zero expectation that we would preserve orthogonality from latent to pixel space. I don't think any part of the GAN objective requires that these interesting features actually be linearized in the latent space (obviously they are not in pixel space), but the approach is to use a GLM to find the latent vectors that best fit the features anyway. Whether or not the vectors you identify with the GLM really retain their semantic meaning through the latent space, they're also clearly not orthogonal, so changing the latent representation along one dimension also changes others.
• colordrops 8y ago
  That's kind of an absurd statement. It's not possible in this universe at all for a human or AI to generate an image of a face without any previous impressions of one.
  • zaroth 8y ago
    If it had any "impression" of what a face was, you would never see the results that occurred in [1]. My point is the algorithm has no impression of what a "face" is at all. It seems like this was given prototype pictures with scoring attributes, and mapped those into some sort of gradient, completely independent of any notion of "face".
  • visarga 8y ago
    The generator never sees a single image. But it receives gradients from the discriminator teaching it to make better fakes. The discriminator here acts like a 'learned loss function' for the generator.
• visarga 8y ago
  > It does not appear to me that we are seeing an AI that has learned what humans look like and can now generate arbitrary fictional humans Here, take a look at this video, one hour long of high-res generated faces. Can you still say AI hasn't learned to generate arbitrary identities? https://www.youtube.com/watch?v=36lE9tV9vm0&t=262s https://www.youtube.com/watch?v=36lE9tV9vm0&t=262s