- Energy based framework makes for more robust models
- An interesting bit: their model internally learns an auto-encoder. But: it’s allowed to reconstruct total garbage for anything that’s not a valid image.
- An exhaustive evaluation shows that EB-GANs are more robust to hyper-parameters than traditional GANs
- They’re also really easy to understand!
I enjoyed this paper. I would have liked if they had compressed/summarized the evaluation section, and provided instead some more background on how their model differs from a traditional GAN, but you can’t have everything.
The authors introduce energy-based generative adversarial networks (EB-GANs) for learning generative/discriminative models. They argue that existing GAN models are difficult to train, as they can easily enter unstable regions. The authors argue that the structure of EB-GANs makes them more robust to hyper-parameter settings. This is primarily based on an argument (that I admittedly haven’t fully grasped) which shows that the decision problem faced by traditional GANs can be framed in an energy context. In this context, the GAN is forced to learn a decision function with effectively an infinite margin, which the authors conjecture makes them difficult to train. There’s a bit of hand-wavyness about this whole argument, but the results of the paper seem to bear it out.
The evaluation is exhaustive (and exhausting to read through), but the net result is clear: EB-GANs much more reliably achieve good training results.