Battling Toward Truth

Published by admin on

Generative Adversarial Networks have features of legal systems of some countries. There is a decision process (the “law”) that reaches a judgement on conflicting interests. The law evolves to address perceived deficiency, and adversaries devise cleverer ways around it.

For example, let’s say you write a program (the law) that detects if there is a cat in an image. Then your adversary writes one to generate images with no cats, but yet fool your program. Iterate by adding the images that fool the detector to its training set with the “no cat” label, and then train the fake cat injector on the retrained cat detector. Rinse and repeat. Now detector and adversary can be left to battle it out and hopefully, they improve.

Shi et al. in 2017 leverage this to make better road maps. The article reference is :doi:10.1109/ACCESS.2017.2773142
https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8106771

They wrote a generator that makes roadmaps from overhead imagery and a discriminator that compares the ground truth roadmap to the generated one. The idea is that eventually the generator gets better. And it does.

I attach Fig 5 showing comparative performance against other methods. The bottom row is the Shi result.

Categories: DataScience