Get the latest tech news
Implementing the Goodfellow GANs paper
ve Adversarial Networks (GANs), discovered by Ian Goodfellow in 2014, were an early method in the area of generative AI. I will focus on image generation as set out in the paper Generative Adversarial Nets1.
Cross Entropy Loss is a good metric for classification problems and when you implement different papers in the deep learning space you’ll come across it alot. The change from input size 1 to 100 is another choice driven by empirical evidence, I’m not entirely sure why it works but my intuition is that as the task is more complex the higher dimensionality aids learning. A more pressing issue can occur where the generated images all look bad and do not seem to improve, when this happens the best solution is to reinitialise the networks and run the training loop again.
Or read this on Hacker News