Challenges

GANs provide an alternative method of developing generative models. Their design inherently helps in mitigating the issues we discussed with some of the other techniques. However, GANs are not free from their own set of issues. The choice to develop models using concepts of game theory is fascinating yet difficult to control. We have two agents/models trying to optimize opposing objectives, which can lead to all sorts of issues. Some of the most common challenges associated with GANs are as follows.

Training instability

GANs play a minimax game with opposing objectives. No wonder this leads to oscillating losses for generator and discriminator models across batches. A GAN setup that is training well will typically have a higher variation in losses initially but, eventually, it will stabilize and so will the loss of the two competing models. Yet it is very common for GANs (especially vanilla GANs) to spiral out of control. It is difficult to determine when to stop...