MIT6036L01b 1

it's tempting there's this a kind of romantic idea about machine learning which is that it's this magic box and you pour data in one end and outcomes like super awesome programs to do things and it's not so easy right and and so and the human has a big role to play in setting up a machine learning program right our machine learning solution to a problem and we're gonna be talking about many aspects of this not all the aspects but I really want to emphasize what you have to do to get this to work right so what do you have to do to make a machine learning application well you have to somehow you have to get data or so that's not always easy there can be all kinds of subtleties about that and and process it in a way that a machine learning algorithm can deal with it you have to think about a space of possible solutions so what what what what could possibly if we're gonna use a computer program to find an answer to a question we do have to at least think about what is the space of possible answers we have to characterize what makes a good solution sort of the objective right so how could I tell whether one solution is better than another we have to decide that to we'll talk about these things in much more detail today I should also say there will be detailed notes posted they're posted already and so don't feel the need to slavishly write down stuff I write on the board cuz it really is all written somewhere else you're welcome to but just just saying that so if space possible solutions figure out an objective what makes one solution better than another one come up with an algorithm so you took someone with some kind of an algorithm to actually find a good solution then we have to run it and we have to validate the results really the only part of this process that the computer does for us completely on its own is this right so we have to approach the problem understand it well enough get the data do all these steps we run it out of order we figure out what happens so we'll think about how how we do all these things okay so when you get data what do you do with it right so one of the interesting problems involved in going from data to some kind of prediction so one aspect of machine learning is actually the part of it that's been studied by statisticians for years and is very important so one part of the problem that we have to solve is estimation right so maybe I'm interested in understanding I don't know what the temperature in this room is and I take a bunch of measurements right and I get one that says sixty three point two and another one that says seventy four point one another one that says sixty nine point zero and I so I get all these measurements and I have to aggregate that data in some way that gives me I don't know some estimate of the temperature in the room and maybe some estimate of my own certainty of what's going on right so I might estimate I might look at the variance I might use the variance of this data to estimate my uncertainty about the actual mean so those are the kinds of questions that get answered in statistics a lot and these kinds of statistical questions are important to us in machine learning but the part that we really actually most focus on is the problem the generalization all right so the problem of generalization might be I measured the temperature in this room for the last three days or the last three months of the last three years and I would like to make a prediction now about the temperature in this room tomorrow and making a prediction about the temperature in this room tomorrow is an act of hubris right it's an act of like saying you know you're asking me to make prediction about a situation that I don't have any data about directly I have data about related situations I have data about what it was like in this room yesterday and two weeks ago and so on but I'm gonna have to use that data now to answer a question that's a brand-new question a question about which I have no examples in particular so generalization is about going from data that's about situations questions problems cases that that are not the same as the one that you want to make the prediction about and so that's like super hard what what even gives us a license to do that so there's a lot of work in philosophy about generalization and induction and so on a lot of work and statistics trying to think about that but this is going to be the thing that we always have to keep in mind is that we're getting data about one thing and we're using it to make predictions about other things in general so that's hard and fun and interesting