ad-free
Channel / Source:
TEDx Talks
Title: The wonderful and terrifying implications of computers that can learn | Jeremy Howard | TEDxBrussels
Published: 2014-12-06
Source: https://www.youtube.com/watch?v=xx310zM3tLs
well it used to be that if you wanted to get a computer to do something new you would have to program it now programming for those of you who that haven't done it yourself requires laying out in excruciating detail every single step that you want the computer to achieve to turn in order to achieve your goal now if you want to do something you don't know
how to do yourself then this is going to be a great challenge so this is the challenge faced by this man office Samuel in nineteen fifty six he wanted to get this computer to be able to beat him at checkers how can you write a program layout excruciating detail how to be better than your check so he came up with an idea here at the computer
play against itself thousands of times and learn how to play checkers and indeed it worked and in fact by nineteen sixty two this computer had beaten the Connecticut state champion so officer was the father of machine learning and I have a great debt to him because I am a machine learning practitioner I was the president of cattle a community of over two hundred thousand machine learning
practitioners capital puts up competitions through trying to get them to solve previously unsolved problems and it's been successful hundreds of times so from this vantage point I was able to find out what about what machine learning can do in the past can do today and what it could do in the future perhaps the first big success of machine learning commercially with Google Google show that it's
possible to find information by using a computer algorithm this algorithm is based on machine learning since that time there's been many commissions successes of machine learning companies like Amazon and Netflix is machine learning to suggest products that you might like to buy movies that you might like to watch sometimes it's almost creepy companies like linkedin and Facebook sometimes we'll tell you about who your friends might
be and you have no idea how it did it and it's just because it's using the power of machine learning these algorithms that have learned how to do this from data rather than being programmed by hand this is also how IBM was successful in getting Watson to beat the two world champions of jeopardy answering incredibly subtle and complex questions like this one this is also why
we are now able to see the first self driving because if you want to be able to tell the difference between say a tree and a pedestrian well that's pretty important we don't know how to write those programs by hand but with machine learning this is now possible and in fact this car has driven over million miles without any accidents on regular votes so we not
now know that computers can land and computers can learn to do things that we actually sometimes don't know how to ourselves well maybe can do them better than us one of the most amazing examples I've seen of machine learning happened on a project that I ran a cattle where a team run %HESITATION black ichor Jeffrey hidden from the university of Toronto won a competition for automatic
drug discovery that what was extraordinary here is not just that they beat all of the algorithms developed by mac all the international academic community but nobody on the team had any background in chemistry or biology a life sciences and they did it in two weeks how did they do this the Houston extraordinary algorithm court declining also important was this in effect for success was covered in
the New York times in a front page article a few weeks later this is Jeffrey Hinton here on the left hand side deep learning is an algorithm inspired to help the human brain works and as a result it's an algorithm which has no theoretical limitations on what it can do the more data you give it and the more computation time you give it the better it
gets the New York times also showed in this article another extraordinary result of declining which I'm gonna show you now it shows that computers can listen and understand now the last step but I want to be able to take in this process is to actually speak to you in Chinese now the key thing there is we've been able to take a large amount of information from
many Chinese speakers and produces a text to speech system that takes Chinese tax and converts it into Chinese language and then we take it hour or so of my own in reviews dot to modulate the standard Texas speech system so that it would sound like me we get the results are perfect there are in fact quite a few areas June moving bulan made would like to
there is much work to be done in this area federal news goals and full moons of Godzilla so that wasn't a machine learning conference China a it's not often actually of academic conferences that you do hear spontaneous applause are although of course sometimes the topics conferences feel free up everything you saw there was happening with the placing a kicker of the transcription in English with the
planning the translation to Chinese in the text right the planning and the construction of the voice was deep planning as well of so deep learning gets to sixty ordinary thing it's a single algorithm that consume to do almost anything and I discovered that a year earlier and it also went to see in this obscure competition from Germany called the German traffic sign benchmark declining had lent
to recognize traffic signs like this one not only could recognize the traffic signs better than any other algorithm the leaderboard actually showed it was better than people about prices for these people so by two thousand eleven we had the first example of computers that can see better than people since that time a lot has happened in two thousand and twelve Google announced that they had a
declining algorithm what future videos and crunch the data on sixteen thousand computers for a month and the computer independently learn about concepts such as people and cats just by watching the videos it's much like the way that humans land humans don't learn by being told what they see but by letting for themselves what these things are also in two thousand twelve Jeffrey Hinton who we saw
earlier out one is very popular image that competition looking to try to figure out from one and a half million images what they pictures of as of two thousand and fourteen when now down to a six percent error rate in image recognition this is better than people again so machines really are doing extraordinary good job of this and is now being used in industry for example
Google announced it last year that they had met every single location in France in two hours and the way they did it was that they fade street few images into a declining algorithm to recognize and read street numbers imagine how long it would have taken before dozens of people many years it's also happening in China by two is kind of the the Chinese Google I guess
and what you see here on the top left is an example of a picture that I uploaded to buy too steep learning system and underneath you can see that the system has understood what that picture is and found similar images the similar images actually have similar backgrounds similar directions of the faces %HESITATION even some with it hung out this is not clearly looking at the text
of a web page all I uploaded wasn't image so we now have computers which really understand what they see and can therefore search databases of hundreds of millions of images in real time so what does that mean now that computers can see well it's not just computers can say in fact declining is done more often that complex new one sentences like this one and now understandable
with declining algorithms as you can see here the Stanford based system sharing the red dot at the top has figured out that this sentence is expressing negative sentiment declining now in fact is near human performance and understanding what sentences are about and what it is saying about those things also declining has been used to be Chinese again at about native Chinese speaker level of this algorithm
developed out of Switzerland by people none of whom speak or understand any Chinese as I say using declining is about the best system in the world for this are even compared to native human understanding this is a system that we put together at my company which shows putting all this stuff together these are pictures which have no text attached and as I'm typing in here sentences
in real time it's understanding these pictures and figuring out what they're about and finding pictures that is similar to the text that I'm writing so you can see it's actually understanding my sentences actually understanding these pictures I know that you've seen something like this on Google where you can types it in things it'll show you pictures but actually what it's doing is it's searching the web
page for the text that's very different tracks the understanding the images it's something that computers have only been able to do for the first time in the last few months so we can see now the computers could not only see but they can also rate and of course we've shown that they can understand what they hear perhaps not surprising now that I'm gonna tell you they
can write up here is some text that I generated using a deep learning algorithm yesterday and here's some text of an algorithm out of Stanford generated each of the sentences was generated by a declining algorithm to describe each of those pictures this algorithm before has never seen a man in a lecture playing a guitar it's seen the man before it seemed like a floored scenic atop
before it hit his independently generated this novel description of his picture we're still not quite human performance here but we're close in tests humans prefer of the computer generated caption one out of four times of the system is now only two weeks old so probably within the next year the computer algorithm will be well past you performance at the rate things are going so computers can
also write so we put all this together at least a very exciting opportunities for example in medicine are a team in Boston announced that they had discovered a dozens of new clinically relevant features of chambers which helped boxes make a prognosis of a cancer a very similarly %HESITATION in Stanford a group there announced that knocking at tissues under magnification they've developed machine learning based system which
in fact is better than human pathologists at predicting a survival rates for cancer sufferers and both of these cases not only to predictions more accurate but they generated near insightful science in the radiology case they were new clinical indicators that humans can understand in this pathology case it actually discovered the computer system discovered that the sales around for cancer are as important as the so cancer
cells themselves in making a diagnosis this is the opposite of what pathologists had been taught for decades in each of those two cases there were systems developed by a combination of medical experts and machine learning experts but as of last year were now beyond that to this is an example of identifying cancerous areas under %HESITATION of human tissue under a microscope of the system being shown
here can identify those areas more accurately or about as accurately as human pathologists but was built entirely with day planning using no medical expertise by people who have no background in the field similarly here this Nerone segmentation of we can now segment near owns about as accurately as humans can but this system was developed with deep learning using people with no previous background in medicine so
myself as somebody with no previous background in medicine I seem to be entirely well qualified to start a new medical company %HESITATION which I did arms %HESITATION I was kind of terrified of doing it but the theory seems to suggest that it won't be possible to do very useful medicine with a using just the state analytic techniques and thankfully the feedback it's been fantastic not just
from the media but for the medical community who have been very supportive the theory is that we can take the middle part of the medical process and turn that into data analysis as much as possible leaving doctors to do what they're best at I want to give you an example it now takes us about fifteen minutes to generate a new medical diagnostic test and I'll show
you that in real time now but of compressed it down to three minutes by cutting some pieces out rather than showing a showing you creating a medical diagnostic test I'm gonna show you a diagnostic test of caught images because that's something we can all understand so here with starting with about one and a half million caught images and I want to create something that can split
them into the angle of the photo that's being taken so these images are entirely unlabeled so I have to start from scratch without declining algorithm it can automatically identify areas of structure in these images so the nice thing is that the human and the computer can now work together so the human as you can see here is telling the computer about areas of interest which it
it once the computer then to try and use to improve its algorithm now they think leading systems actually are in sixteen thousand dimensional space so you can see here the computer rotating this through that space trying to find new areas of structure and it also successfully the human who's driving it can then point out the areas that interesting so here the computer successfully found areas of
for example %HESITATION I'm angles so as we go through this process we gradually turn the computer more and more about the kinds of structures were looking for you can mention in the diagnostic tests this would be a pathologist identifying areas of Bartoszyce for example or a radiologist that indicating %HESITATION potentially troublesome Nigel's and sometimes it can be difficult the algorithm in this case it's got kind
of confused the fronts in the backs of the cards are all mixed up so here we have a bit more careful manually selecting these %HESITATION fronts as opposed to the backs and then telling the computer that this is a type of %HESITATION group that we're interested in so we do that for awhile we skipper for little bit and then we tried a machine learning algorithm based
on these couple of hundred things and we hope it's got a lot better you can see it's now started to fade some of his pictures out showing us that it already is recognizing how to understand some of these itself we can then use this concept of similar images and using similar images you can now see the computer at this point is able to entirely fine just
the fronts of because and so at this point the human can tell that the computer okay yes you've done a good job of that %HESITATION sometimes of course even at this point it's still difficult to separate out groups bar in this case even after we let the computer trying to take this for a while we still find that the left side and the right sides pictures
are all mixed up together so we can attend if the computer some hints and said okay trying find objection that separates out the left sides and the right sides as much as possible illusionistic mining algorithm and giving up that hint up okay it's been successful it's managed to find a way of thinking about these objects that separated out these together so you get the idea here
%HESITATION this is a case not wear their computers being replaced by machines as are the computers of the human is being replaced by computer but where they're working together at what we're doing here is replacing something that used to take a tame five or six people about seven years %HESITATION and replacing it with something that takes fifteen minutes for one person acting alone so this process