3 ms·
Why humans learn faster than AI–for now
- YeGoblynQueenne 9y agoThis is a very nice idea! To use the time taken to complete the game with nonsense textures as a proxy for the importance of background knowledge. However, I can't avoid bitching a bit about how the fact that a strong prior can really boost sample efficiency is really, really old news. That is one of the main ideas in Inductive Logic Programming, where priors are specifically encoded as background knowledge, or a strong language bias, etc. It's a bit disheartening to see people reinventing the wheel instead of looking a little outside their narrow field.
- rach0012 9y agoHi, I am one of the authors of the paper. Thanks for your comment! In our paper, we never make a claim that we are the first ones to show that prior information matters. The goal of our paper is to study what kinds of prior information are more critical during human gameplay. A major finding of our work is that general priors about objects which we learn when are as little as 2-3 months old are some of the most important prior knowledge that guides efficient gameplay. Unfortunately, the media articles have missed this important distinction so it appears as if we are claiming that prior knowledge matters (which in fact the field has known for a long time). Thanks again for your comment!
- YeGoblynQueenne 9y agoHi and Thanks for replying! I did warn I was going to be "bitching" about that bit :) I read your paper- I agree that you didn't make any claims that you're the first to notice the effect of strong priors on sample efficiency. Like I say above, I think your paper is really interesting and that your experiments are a great idea- it's a little disappointing to see it did not get as much interest on hacker news as other recent work on deep reinforcement learning posted on here (but then, this is a popular forum, not a scholarly journal :). One thing I felt was missing (unless I'm the one who missed it) is some examples of what happens when a human player has played the original game once, before attempting the ones with the various ablations. I tried that- I played the un-masked game once, then tried the "hard mode" game, the one with visual similarities masked, and I was able to finish it without dying, although I kept falling off platforms (I guess I got lucky and missed the fire pit). I even managed to kill the blobs. This was all done from memory, of course, but that's memory of object affordances, identities and semantics, not specific keypresses. I think that's an interesting thing to note. But, you know- just a minor detail :) Thanks for writing the paper and making it available on arxiv.