Review Generative Adversarial Nets (GAN)

Paper Title	Generative Adversarial Nets
Date	2014-06
Link	https://arxiv.org/pdf/1703.10593.pdf

Paper summary

Reading notes

Spontaneous Questions

Q1: What’s the details of the training? A1: Several changes in the implementation such as using L2 instead of binary log loss, PatchGAN instead of GAN, etc.

Paper Review

Short Summary

The paper describes CycleGAN, an image-to-image translation approach that works very well even when there is no paired training data. In particular, CycleGAN make the translation a 2-way-street by training both generative model G and its inverse F. Beside from adversarial losses for both G and F, an additional cycle loss was employed to test the models’ ability to reconstruct the exact original image from a translation. The paper applied several techniques such as PatchGANs and L2 loss to stabilize training. It also produces quite comprehensive study of the results against the most recently developed techniques.

Strengths

• The use of inverse function F and a cycle loss is a novel and was proven effective in producing high-quality results
• The study of results is comprehensive thus provide credibility to the model’s claims
• The ablation study helps provide some insights into how the model perform under different loss functions
• Include a study with real human’s perception, which is insightful to estimate performance if applied to a real-world setting.
• Code was released, which is helpful for re-creation

Weaknesses

• Doesn’t have as good performance as pix2pix in many tasks, a contemporary method
• Perception study doesn’t include the performance of pix2pix
• The model seems to perform best on style transfer such as changing color and textual, fails to tackle more complex cases.

Reflection

• It seems like using exact-matching for Cycle loss (L1 in this case) is not the best choice to build a robust model.
• This model is definitely not production-ready, one thing that seems to improve a lot in later models such as diffusions

Most interesting thought/idea from reading this paper

Cycle loss is a nice idea and could be explored further. Combining with more recent techniques such as diffusions can be a good approach.