Jeremy Howard at Davos: Jobs For The Machines
00:00:00.000 | It doesn't know where it's going so last month you might have noticed something Google actually announced that they had met the exact
00:00:08.600 | Location of every single residence and every single business in the entirety of France every single house number
00:00:15.400 | Now if you were tasked with that job
00:00:18.760 | What would you need me a hundred people for a year?
00:00:22.600 | What did Google do Google used an algorithm which did that entire process in one hour?
00:00:30.280 | They met every single residence and business exact location in one hour. How did they do it?
00:00:36.880 | well, they started with Google Street View which is hundreds of millions of images and
00:00:41.520 | They used a machine learning algorithm
00:00:44.320 | Machine learning algorithms are algorithms that you don't really program in the traditional sense
00:00:50.480 | But you just give it a sense of what you're trying to do and it figures it out
00:00:54.400 | so in this case they took a few hundred street numbers and kind of circled them and
00:00:59.200 | Said to the algorithm look for stuff like this and read them the algorithm did the rest
00:01:05.240 | Very similar with this with this project looking for these areas. It's got green arrows are areas that are called mitosis
00:01:13.780 | These are the most terrifying parts of the breast cancer tumor
00:01:18.600 | These were hand marked up by a team of four expert pathologists working in unison
00:01:24.520 | That's because actually doing this is very very difficult to any two pathologists on the team only agreed with each other 50% of the time
00:01:31.800 | they then took this expert assessment for thousands of breast cancer images and they sent it to a
00:01:39.600 | machine learning algorithm
00:01:42.080 | the machine learning algorithm then automatically identified mitosis and thousands more images and it had a
00:01:49.320 | 60% overlap with the shared committee view so
00:01:52.860 | Algorithm is better than any expert pathologist in the world and identifying this incredibly difficult thing of ptosis
00:02:00.440 | This is actually happening again and again
00:02:04.240 | I was involved in a project a couple of years ago for Merck and they wanted to know can we find
00:02:11.080 | Molecules ahead of time which are likely to be toxic and you probably know the punchline
00:02:15.680 | The team that I was involved with did not try and program that but instead gave examples of a whole lot of toxic molecules
00:02:23.440 | to a machine learning algorithm
00:02:25.600 | It took only two weeks for the whole project and at the end of it
00:02:29.440 | Was a system that was much better than Merck and much better than anybody in academia or industry had ever developed for identifying
00:02:36.040 | toxic molecules
00:02:38.800 | I'll tell you something else in each one of these systems
00:02:42.100 | They were developed by people with no previous background in this area
00:02:46.160 | So for example this in the Merck
00:02:48.240 | example both developed by machine learning experts who had no previous background in biology and
00:02:54.400 | Generally these things take days or maybe a couple of weeks
00:02:58.260 | It's not just in medical and imaging areas
00:03:02.720 | This is also being used in language in fact on your Android phone when you're using
00:03:09.040 | Voice recognition it's using this kind of machine learning system. And in fact, I'll tell you something else
00:03:14.600 | Not only are all these best in class algorithms using machine learning rather than traditional programming
00:03:20.640 | They're using the exact same algorithm one algorithm
00:03:25.200 | It's called deep learning and it's actually based on the computing process that the human brain uses
00:03:30.960 | We're now at a point whether the hardware the computer hardware can
00:03:35.280 | Approximate that well enough that we are seeing the best in class performance other examples the world's best algorithm for
00:03:42.680 | Translating English speech into Mandarin speech in the original inflection of the original speakers voice again is a deep learning algorithm
00:03:52.240 | So we're at a point now where a single algorithm, which is actually getting multiplicatively better every year
00:04:00.200 | The more data you give it the more computing power you give it the better
00:04:03.160 | It gets is now at or better than human performance at the very things
00:04:08.720 | Which humans spend most of their time doing all the blue areas here are countries where over 80% of?
00:04:17.360 | employment is in services
00:04:19.760 | More specifically for example in the US
00:04:24.040 | 55% of employment is actually an information processing. This is the exact thing that these algorithms are now at or better than human performance at and
00:04:33.520 | getting multiplicatively better every year
00:04:36.240 | So there are opportunities for example for your organizations these same algorithms are actually also the best in the world at identifying churn
00:04:44.320 | credit scoring at the
00:04:46.240 | Prioritization you know any of your organizations that that use these approaches will get ahead those that don't will fall behind
00:04:53.360 | So it has a big commercial opportunity as well
00:04:55.640 | I think about the other side every single organization which is doing this is actually replacing
00:05:01.960 | Human jobs the most common jobs with machines this has happened once before
00:05:07.280 | And the Industrial Revolution
00:05:10.080 | Before that most people were working in manufacturing and agriculture
00:05:13.540 | Suddenly those jobs didn't exist anymore because machines did it better
00:05:18.360 | Did you know that the decades afterwards 80% of workers were much worse off in fact 80% of workers?
00:05:25.800 | Did not have a basic living wage
00:05:28.280 | Incredibly disruptive
00:05:31.560 | What is this going to mean for us when all of the people currently in the services sector of many of them find themselves?
00:05:37.680 | Literally unable to add economic value because the computers are getting more applicatively better every year
00:05:44.760 | So you might be thinking this is all very nice in theory, but you know wouldn't you expect to have seen something in practice by now?
00:05:50.840 | Wouldn't you expect to have already seen that like the the value of human capital is not keeping up with the improvements in productivity?
00:05:58.740 | Well, yes, you would
00:06:01.320 | In fact the value of human capital as expressed in median income has been flat for the last 15 years
00:06:07.560 | And it's now starting to go down at the very same time
00:06:11.200 | That productivity overall in the economy has kept increasing and increasing
00:06:15.540 | So what I wanted to tell you today is be aware of the exciting opportunities for all of your organizations from machine learning
00:06:22.980 | But also be aware of the threats to our socioeconomic systems that will come in the future
00:06:28.320 | And let's start to think about both the opportunities and the threats