AI is human after all
Recently Stanford and Google have found something interesting in their CycleGAN satellite to maps and back.
Source: https://hub.packtpub.com/cyclegan-learns-to-cheat-by-hiding-information-in-generated-images/
When I was young and a student we had physical education. The teacher was a little bit overweight and smnoked. So, when we had to run a few kilometers, the guy didn't go with us. He wasn't in shape and not even interested in getting into shape. So, the run was in a nearby park. The moment, we were oput ofsight, the majority of the runners took a shortcut. Obviously in a smart way, waiting the right amount of time before reappearing and finishing the run. We, nerds, understand that we cannot finish using a shortcut and do better than a world record on that specific distance.
Obviously, the AI doom people will interprete this story a bit differently. AI does something that the implementers did not expect so we should be afraid, very afraid, because we will all be enslaved and/or killed by terminators if we don't stop this threat to society.
What actually happened is less spectacular. The agent embedded information in the maps it produced to be able to match the original image better. It is n o surprise that in life, you need to know exactly what you are doing. And you need to verify that what you are doing is exactly what you want it to do. They obviously found a weakness where the specific algorithm allows shortcuts. Humans use shortcuts and efficiency optimization all day. The brain tries to optimize and save energy. Driving a car is learned and then it is automatic in the sense that it doesn't require your brain to concentrate on shifting gear and looking in mirrors. You needed to concentrate when you were learbning to drive but once it is learned, it is stored. Certain topics or questions you hear for the first time, you will spend energy to think about it and then the resulting answer (opinion) is stored. The same question (or similar) is a shortcut now. Saving energy for the brain. We don't want to think about that same question over and over. There is a similarity in humans and AI, the path of the least effort. So, when we had to run with 20 students in the park, there would always be one or two that didn't take a shortcut. I was one of those idiots that did the full run and finished outside the required time window allowed for youngsters at that age.
Gym class, humans in general and AI are similar it seems. ASIC design is no exception. Today ASIC's are the fastest and most efficient way for any algorithm. That doesn't imply it is economical, that is another discussion. In design and verification of digital circuits, one needs to understand the whole project from start to finsih, see what effects what in the pipeline and see what can be done about this. It is a million dollar project after all. Would you be surprised to hear that ASIC people tend to take shortcuts as well? To save energy of course :-).
Know what you are doing, the parameters that influence it and the results you get. And verify independently that what you get has gone through the expected processing with the expected results at every stage. It seems they did their job on CycleGAN. That is how it should be.