Consider a particular dataset 'X'. Assume that I have trained my deep GAN on it. Now, can we use the learnt features or weights of the discriminator( which is a neural network ) to build a more efficient classifier, than by training neural networks in an end-end fashion.
There has been lot of interest in utilizing the generative modeling abilities of GANs for other auxiliary tasks like classification. If we consider a vanilla GAN, we have generator and discriminator where the former is trying to generate the samples similar to training data and the later is trying to differentiate between samples from generator and samples from the true data distribution (i.e. training data). So, I am not sure, how could one utilize the discriminator to do the classification within the dataset. It is true that the discriminator won't be able to differentiate between true data distribution and approximated distribution without capturing the variations within the dataset and intuitively, an efficient classifier would also try to capture the variation among the dataset for proper classification. So, it's also a research area of transfer learning.
Also, we should note that, extracting features, or as expressed as inference of latent variables in generative modeling terminology, is not possible in vanilla GAN and hence, researchers are trying to fuse inference within GAN setup. One recent work that I came across was ALI (Adversarially Learned Inference). But, your question was more directed towards utilizing the learned weights of discriminator, and in this regard, ALI and GAN are same. It seems, my answer was not so helpful but yeah, good question.
There has been lot of interest in utilizing the generative modeling abilities of GANs for other auxiliary tasks like classification. If we consider a vanilla GAN, we have generator and discriminator where the former is trying to generate the samples similar to training data and the later is trying to differentiate between samples from generator and samples from the true data distribution (i.e. training data). So, I am not sure, how could one utilize the discriminator to do the classification within the dataset. It is true that the discriminator won't be able to differentiate between true data distribution and approximated distribution without capturing the variations within the dataset and intuitively, an efficient classifier would also try to capture the variation among the dataset for proper classification. So, it's also a research area of transfer learning.
Also, we should note that, extracting features, or as expressed as inference of latent variables in generative modeling terminology, is not possible in vanilla GAN and hence, researchers are trying to fuse inference within GAN setup. One recent work that I came across was ALI (Adversarially Learned Inference). But, your question was more directed towards utilizing the learned weights of discriminator, and in this regard, ALI and GAN are same. It seems, my answer was not so helpful but yeah, good question.
.
the discriminator is trained on a very specific task, that is discriminating between "true" and "fake" (that is GAN generated) samples
your classification problem takes place inside the domain of the "true" examples and does not care about "fake" samples
therefore, it is not obvious that the features learnt by the discriminator, useful to define the "true" vs "fake" frontier, will bring useful information to built a frontier inside the "true" area
in other words, there may not be so much to transfer from one problem to the other because the problems are very different
but, as long as it has not be tried, the above is just my not-so-informed opinion !
.
As both @Prasanna and @Fabrice say, Discriminator network in GANs are for identifying "real" v/s "fake" images. However, there is a GAN architecture called Conditional GAN where the discriminator not only check between "real"/"fake" but also whether the sample generated by Generator network belongs to the specified class also. I think you can utilize the discriminator network of a Conditional-GAN for classification purposes by iteratively checking over all the class labels with the image, but it may be a little bit of over kill for classification task!
Relevant Paper: Article Conditional Generative Adversarial Nets
You can check a tensorflow implementation of C-GAN for MNIST dataset here: https://github.com/v-i-s-h/TensorFlow-Sketchbook/tree/master/GAN/03_CGAN
Yeah. You can do that. But dont expect any good features as you are not guiding the discriminator to extract any specific features. So it is ideal to add the classification head after the discriminator. and train end to end.
Typically to extract features, you can use the off the shelf layer of the network. The intuition is that these features are linearly separable because the top layer is just a logistic regression.
For GANs, you can use the features from the discriminator. These features are supposed to give a probability if the input came from the training dataset, "real images".
Check out this paper:
Article Unsupervised Representation Learning with Deep Convolutional...
According to this post you can
https://code.fb.com/ml-applications/a-path-to-unsupervised-learning-through-adversarial-networks/
Discriminator is just a classifier but your question is a bit unclear. You want to use discriminator as a classifier but what do you want to classify?? If your application is real data and generated data then you are good to go, if you have any other data, then you have to retrain the discriminator on new data set assigning it labels. In short you might need to train already trained discriminator end to end
Check out this paper:
Article Image Recognition With Conditional Deep Convolutional Genera...
As mentioned by Fabrice Clerot , intuitively, as the target of the discriminator is to distinguish the generated samples from the real samples, it will just focus on the difference between these two kinds of samples. But what makes sense is the difference between real samples, which is the samples used by the downstream tasks. However, the feature extracted can be factorized into two orthogonal subspaces. The first space contributes to the discriminator task, while the second is free from it. As in most cases, the features used to distinguish the real samples from the generated ones are noise, the second feature space will be noise-free. From this perspective, although the task of discriminator will not focus on the difference between real samples, which are useful for the downstream tasks, the noise-free features contained in the second subspace will work.
Thank you for useful reply. Please help, how to factorized features into two orthogonal subspaces
http://cs231n.stanford.edu/reports/2017/pdfs/325.pdf
- Badges
- Science topic