ad-free
Channel / Source:
TEDx Talks
Published: 2016-06-21
Source: https://www.youtube.com/watch?v=qIZ5PXLVZfo
it'll Google the site which search results how does Amazon choose which products to recommend how does Netflix movies for you and Spotify pick songs how this Facebook choose which updates so you and Twitter which we the answer to all these questions is the same machine learning I just learning is the most important technology in the world today and yet most people are only dimly aware of
it if at all this urgently needs to change because whoever controls machine learning controls the future you need to control what questions are learning about well your life will increasingly be run by the likes of Google Amazon and Facebook and you may or may not like the results and the first step in taking control is understanding what machine learning his and what it does in the
first stage of the information age I computers have programmed by us they did all sorts of amazing things from running a factory and you know managing payroll to playing chess and making Pixar movies but in each case someone had to explain to the computer what to do in painstaking detail programming had to write down al with them would step by step instructions to accomplish the desired
task this was very slow and very expensive and it limited the rate of progress but in the second stage of the information age which we are now entering we no longer have to program computers because they figure out by themselves what to do by learning from data computers programmed themselves that's what machine learning his computers programming themselves the automation of automation Google's computers look at which
search results you click on in the past and figure out which ones you will quicken in the future Netflix looks at which movies you like and predicts which other movies you like and as the data we generate grows exponentially computers become smarter and smarter with no extra work from us machine learning is a revolution that's going to change how we live work and play and it's
just getting started it's not just online either smart phone in your pocket right now is learning about he uses machine learning to understand what you say to Craig your typos %HESITATION privet what you gonna do next and make suggestions it learns your habitual routes from GPS and whether you tend to be late for meetings by comparing it with your calendar you can even learn how you
walk from its accelerometer one of these days your smartphone is going to warn you and call nine one one I think to about to have a heart attack some machine learning could save your life self driving cars wouldn't be possible without machine learning we actually don't know how to program cars to drive themselves they learn by watching people drive investment firms use machine learning to predict
which thought which stocks will go up or down companies use learning algorithms select job applicants so you may owe your job current to future to learning all with these days a third of all marriages start online and the matchmakers are learning algorithms picking potential dates for people based on their profiles so that our children alive today who wouldn't have been born if not from shin learning
but if you ask their parents you know how the the dating sites pair them up out of all the millions of couples they could have chosen to have no idea I'm a machine learning researcher I invents learning all ones for a living and I've been doing it for twenty years but something is different now knowledge of machine learning no longer belongs just in the lap there's
too much at stake if machine learning can determine our fates individually and collectively then we all need to have a handle on it it's like driving a car only engineers and mechanics need to know how the engine works but everyone needs to understand where the steering wheel and pedals are and how to use them so his machine learning in a nutshell Earhart's the learning album is
just a computer scientist not a scientist who studies computers like me but a computer doing science the computer looks at data formulates hypothesis to explain it test this hypothesis against more data refines them or discard them so one until it's closer than that it has a good theory of the phenomenon that it's starting whether it's the workings of a living cell what people's tastes in music
she learning is really just the scientific method that work except it's being carried out by computers instead of by human scientists now learning almonds are not yet a smart this human scientists of course but on the other hand become look at vastly more data and discover orders of magnitude more knowledge orders of magnitude faster then any human scientist ever could she learning is the scientific method
on steroids there's a new source of knowledge on earth computers learning from data and soon the knowledge from this new source World War the knowledge that scientists accumulated over centuries and that the whole human race accumulated over millennia a learning algorithm is different from the traditional learning all living in one essential way with the traditional winning out over them you need a different al with and
for each different thing you want the computer to do if you want the computer to play chess you have to program it to play chess if you want the computer to do medical diagnosis you have to program it to the medical diagnosis but a single learning algorithm can do an infinite variety of different things pending on the data that it learns from if there is a
chess games it wants to play chess if the dentist patient records and the corresponding diagnoses it learns to make the diagnosis this is extraordinarily powerful and machine learning often can be very simple and yet you can accomplish very complex things even things that we don't know how to program a computer to do like driving a car the learning album then is a master al with it
makes other albums and the holy grail of machine learning research is is to invent the Ultimates called with them one that is capable of learning anything from data we've been pursuing this goal for several decades now and we're getting pretty close there are five main paradigms in machine learning each one of them is inspired by ideas from a different field evolution neuroscience psychology philosophy statistics each
one of them is fascinating in its own right so let's see what they're all about the most obvious way to create a universal link on with them emulate the one inside your skull everything you know everything you've ever learns is encoded in the connections between the new ones in your brain connection nests as they called design learning albums based on this idea their leader is Jeff
Kent a psychologist turned computer scientist who splits his time between Google and the university of Toronto he believes that the way our brain Lawrence can be captured by a single learning all with him and he spent the last forty years trying to discover it in fact he tells the story of coming home from work on the very excited saying I did it I figured out how
the brain works and is that a reply old dad night again but after many ups and downs his quest is starting to pay off backpropagation a brain inspired learning all in that he co invented this taking the world by storm rebranded this deep learning its use by the likes of Google Facebook Amazon Microsoft to do everything from recognizing images and speech choosing ads on this %HESITATION
and you know and search results to show you deep mines the start up that Google paid over half the billion dollar sport is essentially a backpropagation show backpropagation is a remarkably simple album in essence it just consists of repeatedly strengthening or weakening the connection between each pair of nuance in order to improve the the accuracy of the predictions of ya'll with the most optimistic of the
connection this believe that backdrop is the master Holly single learning all in that will be capable of learning anything and therefore of ultimately automating on knowledge discovery but the more sober ones admit that back Protestant a far cry from the master al with and the other machine learning camps have very different ideas on how to get there take the evolutionary as they were led by John
Holland from the university of Michigan until his death in August of last year they believe that evolution not the brain is the master al with back propped maybe useful fight for fine tuning the connections between your new runs but evolution made the brain itself not to mention all life on earth hauling started out in the sixties it you know simulating evolution on a computer complete with
populations of competing individuals fitness course sexual reproduction between the the fittest individuals the whole thing by the mid nineties his fault his followers had succeeded any involving devices like radios and amplifiers starting with random piles of parts and the longer well along the way the amassed an impressive pile of patents these days the busy involving real hardware robots with the fittest individuals in each generation programming
three D. printers to produce the next one so if terminator ever comes to pass this may well be how it happens now most machine learning researchers think that imitating biology whether it's evolution or the brain is that the best a very circuitous path to them after all with better to solve the problem from first principles using what we know from computer science logic and statistics Beijing's
believe that the master out with it is based hero this theorem is a mad mathematical rule for updating our degree of belief in the hypothesis when we see new evidence as we see more data the hypothesis that are consistent with it become more likely and the ones that are inconsistent become less likely and to hopefully there's a clear winner Beijing's are the most fanatical of the
five machine learning tribes until recently they were persecuted minority in statistics but these days there on the sand Beijing's believe that if a learning all of them is not consistent with the speed and then it must be wrong but Beijing learning is competition we very expensive and it didn't really pickoff and to who the pearl a professor at UCLA made the breakthrough for which he received
the chewing award the Nobel Prize of computer science at the thousand eleven pearl inventive what are called vision networks a type of model that can very efficiently in code the interactions between thousands or millions of variables your first self driving car will probably have a vision that work inside now for the symbolists the machine learning camp that is closest to classic knowledge base the I. vision
networks are still not powerful enough symbolist like imperial college's Steve Muggleton believe very truly general purpose learning all over them has to be able to freely combined wolf and they discover those rules by filling in the gaps in deductive reasoning if I know that socket this is human what else do I need to know to infer that his mortal that humans are mortal of course and
now I can add this newly discovered rule to my knowledge base eve is this is a robot scientists at the university of Manchester that works on this principle starting with basic knowledge of molecular biology you formulate hypotheses runs love experiments to test them and so one all without human help in two thousand and fourteen is discovered the new malaria drug and now we can make millions
more like even have millions of scientists working on progressing the state of our knowledge know where the al with those of the symbolists emulate the thought processes of a scientist the albums of the analogize there's the fifth and last may to machine learning try I'm more like a lease the child's that doesn't study for the exam and then improvise the answers when faced with a new
patient to diagnose analogy based learners what they do is they just find the patient in their files with the most similar symptoms and they assume that the diagnosis will be the same now this may seem very naive but a knowledge of this have a mathematical proof that he can learn anything given enough data so your mom taught you that procrastination is bad but in machine learning
you can actually be quite powerful Douglas Hofstadter the author of Goodall issue box is a famous analogize and he has no doubt that analogy is the master all with whether or not that's the case learning by analogy has already proved its chops into bhikkhunis recommender systems that recommend products you might wanna buy list on the ones you but before so who will win the race to
invent the ultimate learning algorithm with all the major tech companies pouring resources into it it's hard to predict but maybe none of the tribes has all the pieces of the puzzle and what it will take is a combination of ideas from all of them a grand unified theory of machine learning picking through the standard model physics or the central dogma biology well maybe it'll take an
entirely new insight which could well come not from the professional researcher but from an outsider well from a student in a dorm room like Jeff Hilton was when he started on his quest so if you have such an insight please let me know so I can publish it either way the next decade is going to be one of accelerating change today each company has a little
model if you based on just the sliver of your data that has access to Netflix has a model to predict your movie tastes best when your movie ratings Amazon has a model to predict what you're going to buy based on what you've done on their site and so on but all these little models are quickly coalescing into bigger and bigger ones and soon you have a
complete three hundred sixty degree degree model if you that learns from all the data and the cysts you with everything that you do in your life from buying things and making appointments to finding a job or a mate our digital alter egos will be even more indispensable to us than our smartphones and the world economy will revolve around them our society will become a society of
models everyone's models will be continually collaborating competing negotiating in cyberspace to determine what that happens in the real world you click on the find me a job button on linkedin and your moral instantly interviews for all the open positions that match your specs the same time another copy of your model can be looking for a car for you exhaustively researching all the options and haggling with
the car dealer but so you don't have to if you're looking for a date your model will go on millions of dates with thousands of other people's models and select the most promising wants to try out in the real world but your data and your model have to be under your control not owned by some third party that may have a conflict of interest Sergey Brin
says that Google wants to be the first half of your brain but do you really want part of your brain possibly trying to show you ads probably not we need something different maybe something like data banks that store your data and use it on your behalf in the same way that regular banks store investor money well maybe we need their unions even the balance of power
between us much companies in the same way that labor unions even the balance of power between workers and their bosses and you need to be able to interact with your model setting its goals asking it to justified suggestions telling it where it went wrong and why all very different from the black boxes that we have today and finally as a society we're going to have to
decide what kind of society of models we want to what's so loud what's not how do we make sure that everyone benefits how do we smooth the transition there is lots to figure out if we do there's a bright future where our lives will be happier and more productive if we don't it'll be a huge missed opportunity it's in our hands thank you
