GAN Weekly update, week #29
Hello GAN enthusiasts!
This week, I will summarise three papers that got my attention in the Generative Adversarial Networks (GAN) research area. Here they are.
- Partially Conditioned Generative Adversarial Networks (arxiv.org link)
- InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning (arxiv.org link)
- Adversarially-Trained Deep Nets Transfer Better (arxiv.org link)
Partially Conditioned Generative Adversarial Networks (arxiv.org link)
This paper provides a new method for the conditional part of the conditional GAN: partially conditional GAN. Basically, the authors argue that while conditioning GAN on a dataset, we may not know the whole labels of the data set but partial labels, and if we provide partial information to conditional GAN it works better than classical conditional GAN. Here is what the paper means by saying partial information. The picture below demonstrates the labels of a dataset, which are not complete. For the samples of number 9, you can see that there are three labes for the digit 9.
And here is the result of comparison between conditional GAN (CGAN) and partially conditional GAN (PCGAN).
It looks promising! I highly suggest you check the paper!
InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning (arxiv.org link)
Here is a great paper! This paper focuses on the well-known problems of GAN training: mode collapse of the generator and forgetting of discriminator, and the authors argue that they solve this problem by providing a new framework that can wrap the target GAN that is subject to train, in such that there is no need to change any parameters in the target GAN. This enables researches focus on their research for their particular GAN architecture and forget about the training mode collapse and forgetting of discriminator, which is great!
The authors tested the newly proposed framework with five datasets and showed that their approach significantly stabilizes GAN training and improve GAN performance as well!
Here is the architecture of newly proposed framework, namely InfoMax-GAN framework. In the figure below, notice that blue glyphs are the target GAN and the green glyphs belongs to the newly proposed framework. You can see that the target GAN is not changed but its outputs are fed into a new neural network architecture (they are called critics networks by the authors).
The figure below compares the accuracies of the original GAN and infomax-GAN for the image dataset CIFAR-10. It is very obvious that how Infomax-GAN stabilizes the training of target GAN.
I highly suggest the readers check the paper!
Adversarially-Trained Deep Nets Transfer Better (arxiv.org link)
This paper brings a new approach to transfer learning, a powerful method to train already-trained neural networks via Generative Adversarial Networks. The authors argue that the newly proposed method produce better results if the new task is lack of enough labels or weakly labeled. I suggest you read the paper; it is cool!
In the paper, the authors show that introducing a bias with GAN to the inner layers results in more natural image representation, thus, better transfer learning. To prove their argument, the authors conducted extensive number of experiments on multiple datasets with varying numbers of fine-tuned convolutional blocks and subsets from the dataset of new task. In the paper, the authors provide the model accuracies on the target data set, which proves that adversarially-trained models have outstanding transfer learning characteristics across all configurations.
The method that the paper is proposing for adversarially-trained models is quite different from the other learning methods. The method first tries to maximize the loss by running masks on each image and the try to minimize the loss function. However, the minimization of the loss is not quite easy and the authors provides projected gradient descent (PGD) for a parametrized epoch for example k=20. I highly suggest reader to check to method and try to understand how their new method makes transfer-learning better. Here is the figure that shows the result of the method.
Please focus on the right plot(b). The figure-b shows the test accuracy delta (not test accuracy!) of the new method on different image datasets. According to the plot on the right hand side, fewer number of training images of the adversarially-trained fine-tuned model still provides accuracies very close to the original training model, and if the number of the training images increases the accuracy becomes very close to the original model, i.e. test accuracy delta becomes zero.
This method is very good if you have limited or weakly labeled dataset! I highly suggest researcher read this paper!
I guess this week is again very promising, There are three great papers and each brings a different perspective to the GAN research area. so cool!
