[tech] active learning

the theory lunch today was about active learning theory. the talk was given by sanjoy dasgupta. active learning is an interesting idea, and probably quite relevant to what's going on in the real world wrt learning. the idea is this: given an unlabeled set of points from some distribution, you are allowed to ask for the label for any point, but asking for a label has a cost. the goal is generate an hypothesis that correctly categorizes the points into their respective label categories.

so for instance, if you have points along a line, and say all points above 5 are labeled +, and all below five are labeled -, then a good hypothesis would be to say: "label all points above 5 as +, and all below as -". a not so good hypothesis might be to say: "label all points above 4 as +, and all below as -". but you only see a finite number of points, so in any case, you need to generate your hypothesis based on all your unlabelled points you've seen, as well as those points that have a label because you asked for their label. so ask for labels wisely!

something more concrete, suggested by a post-doc (don't remember the name) at the lunch, and modified slightly by me: suppose you have a user interface, and you want to suggest things to your user to help them along. every time you suggest something, you get immediate feedback from the user: was the suggestion helpful or not. now think of user interactions as your unlabelled points, and the user's feedback based on your suggestion as the label. however, notice that there's an implied cost in asking for a label, because if you asked the user confront a suggestion at every possible interaction, the user would get very angry. so the cost is related to your user's happiness. you need to weigh the benefit you can provide to the user by giving them a suggestion vs the annoyance of giving them a suggestion that they don't like.

anyway that was a bit fast and incoherent, but if you're really interested, talk to me or check out the links.