MIT6036L01a 1

let's do machine learning that so here we are this machine learning now unless you like live under a rock you've been reading about machine learning in the newspaper and everything and so the question is like what in the world is it how do we do it that's we're gonna do in this class and the way I think about machine learning especially is that then fundamentally it's about getting data in some form and aggregating it in a way that lets us make some kind of predictions right so it might be that we would like to predict what's going to happen to the stock market or to the weather or a robot might need to predict what would happen if it turns left some kinds of predictions can also be actions right so in the robotics is that that I work in machine learning applied to robotics so I use a lot of robot examples so it might also be that you're gonna make predictions about what would be a good action to take right so there's all different kinds of things you can predict but fundamentally for us it's about taking data and not so much just analyzing it and getting insight from it but taking data and using that data to actually kind of do some job and so the question is well you have to describe what the job is that you want to do and how to measure whether you're doing a good job of it so I what's happened I mean so was she learning as so many things you know it's been the cool thing and then kind of not so good and then cool again and not so good this is like at least the third epoch of coolness of machine learning is it's way cooler than it used to be so that's good and and really what's what's stunning this time around is that it it it works well and reliably in a large set of applications and basically now for any problem that involves dealing with some kind of signals images or speech or language the only way really that we make those applications anymore is through using machine learning so you know a story that I like to tell is that you know say 20 years ago was maybe the the beginning of this the the of machine learning taking off in image analysis so at that time people were interested in finding faces image right so now everybody's got a camera in their pocket and the camera in their pocket can find the faces in the images no problem with jaws little green boxes around everybody's face so 20 years ago people were trying really hard to do that and lots of very smart people who are trying to write programs that would find faces and images and they would say oh well maybe we should look for eyes and their nose and understand how they should be related lots and lots of really smart people wrote lots did lots of work trying to find faces and images and mostly it didn't work very well and then people turn their attention to the problem of instead of trying for the humans to write the program to do the job the humans started writing a program that would take in data this is an image of a face this is an image of a not face here's another face here's another not face and pour that into some program which the humans still have to write and then out from that would come something that could tell a face from a not face so we're still writing programs and the human and the engineering input is very important but it's sort of now happening at one level of abstraction up instead of trying to write the program to recognize the face we try to write the program that can analyze the data to decide how to recognize the face so so it's been you know enormous ly successful as everybody knows from reading the web