Make Gross Monstrosities/Literal Animorphs with Google's BigGAN!

@Moose Requires google sign-in! Nice try, scammers.

@medz I mean, it was made by Google, but sure.

@phendrick They use pre-trained networks with 1000 image classes and specific politicians are not present there. You’ll need to collect lots of images of one or another and train a GAN on this data set.

@Fduch if you develop the machine to learn, is it really learning or just following your coded instructions?

@medz There are no coded instructions per se.

You just give it a book of problems and punish it for wrong solutions. Then you make an exams with problems that were not shown before. Are good exam results an indication of understanding? It depends.

If your questions are not good enough, it’s easy to cheat and have good grades without needing to understand anything.
What if the student managed to sneak all their notes to the exam? In some cases that makes it possible to get good exam result without much understanding. In that case, we need to limit the amount of information what the ML model can store, forcing it to find a more generalized explanations.

In practical terms, there is a whole field of “explainable” models. You can try changing you test examples a bit and look at the model reaction. Maybe the way it learned to discern between the wolf and a dog is just the presence of snow in wolf images?

I like to compare learning (and also programming) with compression. When you study Math, you’re given a big list of questions and answers (“if question=X1 then answer=Y1; if question=X2 then answer=Y2”) and you can learn a way to compress all that information to a small formula.

A lot of what brain does is just compressing a lifelong live video/audio/senses stream to a very limited storage of human brain. To do that, the brain needs to know how to compress an big image of hamburger to a small concept of hamburger; to compress a full video of a dull work day to a concept of a typical dull work day; to compress a video of a falling leaf to a concept of leaf + knowledge of physics. This conversion is lossy and the brain needs to learn what’s ordinary and what’s unusual, what’s important and what’s not important. Can you remember the shape of yesterday’s clouds in the evening? Probably no. unless they were really special. One important part of video compression is frame prediction.

And being able to predict what’s going to happen is pretty much what we’d call understanding.

@Fduch @medz Your post reminded me of this Google sheet that has been collecting hilarious examples of AI “cheating” at games/tests. Examples include:

• An artificial life simulation where survival required energy but giving birth had no energy cost; one species evolved a sedentary lifestyle that consisted mostly of mating in order to produce new children which could be eaten (or used as mates to produce more edible children)
• A boat that ignores the race entirely and keeps going in a circle hitting the same reward balloons and, despite repeatedly slamming into walls and catching fire, manages to rack up a higher score than otherwise possible
• Creatures bred for speed grow really tall and generate high velocities by falling over
• A program designed to sort lists started deleting everything in the lists because if there was nothing in the set, then it was considered sorted
• Various examples of agents that avoid losing by either crashing the game, pausing it indefinitely, or killing itself at the end of level 1 because it gains more immediate reward than if it progressed to level 2

@Fduch @Moose

@rtjhnstn Not what you were looking for but it’s all you’re gonna get