Hi Everybody
I have a question. I am looking for your opinions and welcome to put any thoughts. In advance, I will appreciate your time and thanks for that. In general, generative adversarial network is able to generate super realistic images by starting from the noise vector. Some times our aim is producing new samples by learning the underlying data distribution. GANs are able to learn such distribution in an indirect fashion. However, most of the times we are trying to add condition to the GANs. In conditional GANs, we are starting from label, image, or etc. For sake of simplicity, let us think about the Super-Resolution task. GANs are starting from low-resolution image and through the learning; they will reach a high-resolution image. My first question is: 1) How GAN add extra information to the low-resolution image? In other words, how it is possible to go from less-information to high-information? Is it simply learn from the historical data? if yes, how it can generalize to the un-seen data? My second question is: 2) through the mentioned process, nature of the data can be potentially changed or not? For example, if our low-resolution cat image has two small eyes, is it possible to have three eyes in generated high-resolution image? In general, How can I conduct systematic experiments to show the conditions that GANs are falling to hallucination mode? Feel free to share your thoughts and opinions. Thanks.
1) Yes, the basis for generation is historical data of sufficient amount. For upscaling LR images to HR images, you are training the generator network of GAN to output HR with LR input. This does not generalize well if you have 1-10 image pairs (LR and HR versions of the same image) to train from, but if you data consists of 1000 to 10 000 image pairs, the generator learns the underlying rule of how to upscale in general. So going from less information to high-information is just an educated guess created by the generator network.
2) The consistency of your generator network depends on the training data that you have. You could do systematic experiments by first characterizing you cat data (10% are brown cats, 90% are white cats etc.), do stratified sampling from that data by taking 100%, 80%, ..., 20% slices and then demonstrate how well the generator network performs with different amounts of training data.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images