(upbeat music) - [Interviewer] A lot of people think of machine learning and they might think of calling customer service and getting that machine that tries to route you to the right person. Can you give an example of a more robust machine learning application, currently in use? - Yeah, so, oftentimes, you know, just a few years ago, the maximum use of AI was finding cat photos on the web, (laughs) but think about today and how AI is being used.
There's obviously the classic ones that people think about, is how people are using it to advertise better, but let me give you an example that people don't think about. We had a team that went in, of data scientists from the university of Chicago, at the fellowship program that takes data scientists and has them work on problems of social good. They went to a police department in the South and what they were focused on is this question of when is an officer likely to be, can you tell if an officer's going to have, use excessive force? And so they had access to all the data, and they were riding along with the officers, and so they started to ask this question of, like, hey, can we use machine learning, AI, against this? And so they start looking for the features, and one of the first things when you're using machine learning is, you get these descriptors that describe, you know, sort of the elements of what the problem is, so a variable in that case might be, oh, this officer has lots of complaints, or this officer has a track record of accidents, or other type of things.
And so we're first to remember, most officers, by and large, are good, hardworking people. There's a signal in the noise problem. There's a small number, given the large number of officers, so, how do you find that? Well, the machine learning finds very quickly the first set of variables are those ones, of complaints, accidents, other than those obvious things, but here's what's interesting, is about halfway down in the feature set, two variables show up, two features show up.
One is the officer responded to domestic violence where a child was present. The second one is, the officer responded to a call of a suicide. So, why those two features, and what's going on here? Well, this case, the data scientists are bolted on with the officers in the ride alongs, so they see it right away. What's happening, they realize, is this officer has responded, the suicide is physically messy, it's a tough emotional environment, same with domestic violence where a child is present, this is very emotional.
Once they're complete with that incident, they may not be the person who's having to write the report or other things, but what does dispatch do? Sends them back out on the street. Go back on patrol. It doesn't do the hard work of saying, hey, you know what, maybe this officer needs some time to decompress. So now if this officer pulls over somebody with a broken taillight and they're flippant, they may not have the right emotional response, and so, in this case, the machine learning could be designed to actually build a smart dispatch system that takes that into account.
And it's crazy to me that we can figure out how to move through traffic (laughs) with using machine learning, to figure out, like, hey, this route is better than this route, but we're not applying it to this problem of how to ensure that the public is taken care of, and the officer is well taken care of, 'cause they need the time to emotionally, you know, decompress, and so this way of thinking about where AI is happening, is not the typical example, in all, there's so many more cases where the untypical examples are going to become easier, and it's not just about finding the solution, it's about, how do you carry that finding into developing a new system, which is the dispatch system in this case.
That's the beginning, and we've just barely begun to apply this in machine learning, AI, for health care, or in drug discovery, or these type of applications. (upbeat music)