Should an AI Learn Like Humans?

Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér. This paper reveals us a fundamental difference between how humans and machines learn. Imagine the following situation: you are given a video game with no instructions, you start playing it, and the only information you get is a line of text when you successfully finished the game. That's it! So far, so good, this is relatively easy to play because the visual cues are quite clear: the pink blob looks like an adversary, and what the spikes do is also self-explanatory. This is easy to understand, so we can finish the game in less than a minute. Easy! Now, let's play this. Whoa! What is happening? Even empty space looks like as if it were a solid tile. I am not sure if I can finish this version of the game, at least not in a minute for sure. So, what is happening here is that some of artwork of the objects has been masked out. As a result, this version of the game is much harder to play for humans. So far, this is hardly surprising, and if that would be it, this wouldn't have been a very scientific experiment. However, this is not the case. So to proceed from this point, we will try to find what makes humans learn so efficiently, but not by changing everything at once, but by trying to change and measure only one variable at a time. How about this version of the game? This is still manageable, since the environment remains the same, only the objects we interact with have been masked. Through trial and error, we can find out the mechanics of the game. What about reversing the semantics? Spikes now became tasty ice cream, and the shiny gold conceals an enemy that eats us. Very apt, I have to say. Again, with this, the problem suddenly became more difficult for humans as we need some trial and error to find out the rules. After putting together several other masking strategies, they measured amount of time, the number of deaths and interactions that were required to finish the level. I will draw your attention mainly to the blue lines which show which variable caused how much degradation in the performance of humans. The main piece of insight is not only that these different visual cues throw off humans, but it tells us variable by variable, and also, by how much. An important insight here is that highlighting important objects and visual consistency are key. So what about the machines? How are learning algorithms affected? These are the baseline results. Adding masked semantics? Barely an issue. Masked object identities? This sounds quite hard, right? Barely an issue. Masked platforms and ladders? Barely an issue. This is a remarkable property of learning algorithms, as they don't only think in terms of visual cues, but in terms of mathematics and probabilities. Removing similarity information throws the machines off a bit, which is understandable, because the same objects may appear as if they were completely different. There is more analysis on this in the paper, so make sure to have a look. So, what are the conclusions here? Humans are remarkably good at reusing knowledge, and reading and understanding visual cues. However, if the visual cues become more cryptic, their performance drastically decreases. When machines start playing the game, at first, they have no idea which character they control, how gravity works or how to defeat enemies, or that keys are required to open doors. However, they learn these tricky problems and games much easier and quicker, because these mind-bending changes, as you remember, are barely an issue. Note that you can play the original and the obfuscated versions on the author's website as well. Also note that we really have only scratched the surface here, the paper contains a lot more insights. So, it is the perfect time to nourish your mind with a paper, so make sure to click it in the video description and give it a read. Thanks for watching and for your generous support, and I'll see you next time!