Cognitive computing | Jerome Pesenti | TEDxBermuda



in 1998 I went to Korea University to study and do some research in artificial intelligence two years after I started a company in that field and two years ago I sold it to IBM now I have been that field of AI of our artificial intelligence for past 16 years yet if you ask me five years ago if I thought computer could do what they do today I would have laughed I would not believe you now what's so amazing what what computers can do today what's this new down of cognitive computing and how did they do that well let me present to you what son from 1911 to 1917 this romantic Russian composed etude tableau for piano Watson who is Rachmaninoff Rachmaninoff is correct and that adds dearly you're a 13400 go again don't worry about it for 1,200 you just need a little more Sun you don't have this hereditary lack of pigment Watson what is albinism good Cambridge for sixteen hundred and thirty daily level what are you gonna wait I'll wager six thousand four hundred thirty-five dollars I won't ask now what son may sound a little silly sometimes but it did go on and win the Daily Double did go on and actually beat by wide margin the best two players and joker in the world and did earn a million dollar for a charity the question may ask is how does Watson do it at first sight you imagine okay what son must have this huge database of knowledge and it takes that question and convert it to our language you can understand it will query that that big base of knowledge and then find the answer actually it doesn't really work like this instead what Watson does it works actually based on a lot of very messy knowledge gets the information from the internet how reliable is that right it will try to crawl that information gets all the text and then from that information it would generate hypotheses candidate answers not just one dozens of them and for each of these candidate answer it will try to create evidences that that answer is actually the right answer next you can actually learn what's the importance of each evidence do they really matter does it tell you that this is going to be the right answer the right question it would rank all these candidates and basically answer was the first one if it's confidence is high enough as its method that's more statistical than just based on on fixed knowledge is not new it was not invented for Watson's there's actually a long tradition of these kind of techniques and you have to go back 40 years again to another IBM researchers to figure out where that comes from well this is Fred jellineck he was an IBM researcher in the 70s and he decided at some point to tackle the problem of speech recognition now the time there was a lot of influence from a big linguist called Nam Chomsky and everybody who was in that film thought that you know the way to crack these linguistic problems was to actually you know look at the language and understand the grammar let's try to try now jellineck was not a linguist it was actually an information theorist you know it will you will deal with signal processing now what he decided to do is to put a team of languages and engineers together to try to crack that problem now the story goes that one day one of these languished quit and jellinek was not able to replace it with in language so he actually replaced him with an engineer not with big surprise a few weeks later he realized the performance of the system actually improve it was better at speech recognition now I decided in an experiment you went to see you next language and say hey why don't you go find a job somewhere else now what happened is that a few weeks later the performances system you know improved again so then one by one he talked to all his linguists and convinced them to go find another job somewhere else and at the end he had no linguist left just engineers and the performance of the system was actually the best of its time now to understand how you can solve a products fixer carnation just with statistics let me jump another 20 years later now at the time it was still the same team and they decided to address another problem which is a little easier to understand which was machine translation not you'd imagine if you had to create a system a computer that understand language to be able to do machine translation you know you will have to understand the grammar the semantics the syntax of one language and try to map that to another language one wrong that's not what they did instead what it uses data they use what you could call a gigantic rosetta stone a set of translation examples of translation you know gigabytes and gigabytes of translations and they will just look at it from an information theory standpoint what they do is look at a few words on one sided you have to translate and Finance its huge corpus what are the most likely translation in these three words then move that one word and look at the next three words and figure out okay what are these three were most likely to be translated in that big corpus but chaining just as little words like this and figure out the most likely path in all this potential path you come up with in translation and once again at the time it became the best the best algorithm to the mission translation to this day when you go to system that do machine translation they use the models designed by that team and when you put the speech and all translation together you get something like this like to reserve a flight from New York to Beijing tonight at 8:20 well computing my pocket our ship Antonio evaluating that hum tight ginger Miyagi Bentham ciao hi shel on Bangkok would you like to book a one-way or roundtrip ticket a rent trip ticket returning on the 23rd of August you don't value it she somehow thought we did now how cool is that I imagine now you could actually look cool without having to fork a thousand bucks for an iPhone iPhone 6 right now you know the best example of statistical learning is actually our brain the brain doesn't learn through rules of fixed knowledge it actually tends to learn by examples and by repetition now in the 50s you know people try to apply you know understand what happens the brain and try to apply it to computers and it didn't really succeed because the computer were not very powerful but in the last five years the huge improvement I want to show you in a minute actually all came from try to mimicking what happens in brain through what we call neural network or deep learning let me give you an example of first tack the first commercial test I was resolved through these artificial neural networks and explained why they work so the problem here is that when you send mail right you write a zip code on an envelope and then you know you could tell the system where to send your letter or way to route it now what if we could do that and recognize the difficult automatically so there's a team of researchers and night C to try to tackle that problem and what the fauna is that it was much harder and we thought because when you look at numbers and all the different ways people have to write these numbers if you're trying to explain you've tried to write some rules that says well this is a 4 this is a 3 so 2 you will always find another person another way of writing the numbers I would defeat that rule instead the creating system using this neural network that can look at very very tiny evidences and lots of evidence back to what I was telling you to you about Watson same principle find a lot of evidence to decide if something is a for something is a tree the three but seven answers I'm not they're very subtle and you can get the network to learn these evidences by training it and through this they actually to create a system that's used today to rot your mail now the beauty of neural network tripe is that not only they're able like Watson before to weigh the different evidences in combine it and make a determination is very subtle but they able even come back with the evidences when you have a network like this and the reason we call it deep is because there are many layers you feel in the raw information the image of the sound and the system will actually learn to extract meaningful attributes automatically out of this you know a diagonal line something that looks like a face or something that looks like a cat face the system can learn to extract very advanced feature and combine them to be pretty to make pretty advanced predictions and what's remarkable about these neural networks is what they have been able to do in the last five years let me go through a few examples now back to speech recognition I believe that everybody in this room not only has a neural network in their brain but they actually have them in that pocket today most of the spiritual cognition system when you talk to your phone and you try to recognize your voice actually are based on neural networks divided by either by IBM Google or Microsoft this is a state of the art today now speech is not yet there it made a lot of improvement is not in there let me show you some other tasks where computer are getting closer to what we can do this is their interesting task it's about recognizing what's on the image so image net is a collection of millions and millions of images I've been tagged by people to say ok this image is about a container ship this image is about a little part or mushroom or cherry now the question is could you a gallon computer when you show it the next image to determine what's the main topic when it makes subject on that image that competition was created in 2010 at the time the best computer could actually have a 30 percent error rate so one out of three would make a big mistake ok now then you're after it was 28% but then people started using this neural net I talked about before deep learning and their error rate went down to 15% the year after it went down to 10% and this year the best thing got 6% now that may not tell you much but a trained expert in that task will actually get a 5 percent error rate so computers are 1% away from what computer far more humans can do what's remarkable about this is that to go from the 15% to the 6% people did not have to invent new algorithm they didn't have to find new ways of determining what's difference between a mushroom and a cherry what's different between a human face and a cat face they actually threw just more data at it and created much bigger Network in a PSD Network store so well that when you throw more data and when they become bigger they become more subtle and can make these very tiny determinations now not only computers can be close to even performance sometimes it can be better and the next one is really ironic on I'm sure you haven't seen these things this CAPTCHA or capture my design so that when you go to a website the website can determine that you are actually a real humans now irony here is that Google actually bought the company bought CAPTCHA but they also have a team that created a system that now can break CAPTCHA better than humans so what's the point right now another task evil even a little more creepy that Facebook that Aetna has access to all your pictures as determine assistant that by feeding a lot of pictures of people is able to recognize the face of people better than human imagine a computer now can recognize your friends better than you I mean getting that now what's remarkable about all these advances is that they all use can is same type of algorithm these neural networks that are very similar to how the brain works now thought right now is that we can actually expand that even more and I'll show you a couple of things that my team is working on today what we believe is the next frontier is to apply this neural networks to language as a previous speaker talked conversation language is very very song very lots of Salty's and it's very hard to try to articulate to formalize to create a language that's so deep that can understand eliciting all these subtleties but we believe that neural networks actually can learn through training to actually identify all these Salty's and understand language as well as human now beyond language what we can do is in your network is combine different modality you know I told you they can handle speech they can handle vision they can handle language how about we could combine them let me give you an example and my team is working on when you're in a room and there's a lot of noise lots of people talking if you talk to someone actually you don't just hear what they say you actually look at their lips moving and when you can look at their lips your comprehension of what they say is actually much higher we're actually able to create system using your own network today we can not only either but also look at the lips of the speaker and get have as much gain in comprehension as humans but if you see where I'm going here is that we are able to create these networks that can understand sound understand vision understand faces understand language and you start looking very much like all these modules of the brain and we believe that we can combine these things and as combining them we will get to something that looks very much like human cognition now the next question you may have is okay so when will we get there one will computer-like pile here on the left would actually match human performance well you know a lot of computer sciences that made the full of themselves by making very big predictions I'm not going to do that but I'll give you some facts here on the left is what we have today so these networks actually are not that easy to reproduce so the biggest that were able to create today have millions of nodes and billions of connection what happens that the brain is actually much more powerful than that actually 100,000 times more powerful it has 100 thousand billion nodes and 100 trillion connection now if you believe in Moore's no which is still working this day which is that computing power increases no doubles every 18 months you get to a number which is in 25 years we should be able to match this now does that mean that we'll be actually able to match all human power I don't know if you ask me I would say no remember that also 16 years ago I didn't think will be here where we are today so I think it's a real possibility that soon in our lifetime we'll see computers becoming as powerful as human but what I bet on it I don't know I don't think so I was sure is that there is this convergence between trying to understand our brains and trying to reproduce through machine learning what they are doing now if you don't believe me that this is you know happening there is something really tremendous in the making here to just follow the money recently Google actually acquired a company called deep nine and it we created by this guide the mrs. Xavier a neuroscientist and all that company actually done that we could we can see today is to create a neural network that could play video games but Google thought it was so it's so good that it spent three hundred million dollars on that company without any product just people now let me finish this presentation with with one life thought on one hand what's amazing about this is that as we are able to create this system that behave like human we feel like we are getting close to this holy grail of really understanding how human work and how human you know how human cognition works but on the other hand the big criticism that these networks and these deep learning techniques I showed you evoke honor is that we actually don't quite understand how they work and remember this little task I show you about trying to figure out the different images if I look at a network and figure out ok how does it network differentiate a human face from a cat's face and a cat's face from a plant on a plant from a stone there's nowhere I can look this is going to show me the features I cannot verbalize it I cannot I cannot rationalize it and maybe at the end what it tells us is that this quest we have had for hundreds of year you know of language trying to formalize language of psychologists trying to understand you know how we behave and formalize it maybe that quest a little bit to tile in the way because what is there to understand more than this neural network maybe all we are and all it is is this really tremendously parallelized this fantastic you know adaptable changing ball complex neural networks and amazing learning thank you you

16 Comments

  1. Aline Bora said:

    Very good and informative talk. I think he has a lisp though.

    June 27, 2019
    Reply
  2. Ernest Kapesa said:

    love the way you talk ūüĎŹ

    June 27, 2019
    Reply
  3. p80mod said:

    Yeah, replace all linguists with engineers, and the world will be better.

    June 27, 2019
    Reply
  4. theUglyManowar said:

    yawn! ( not really… actually amazing BUT ) do not see neural or quantum being pushed in any amazing direction? A machine can do something a human can't? Like what? Go faster like my car? Or toast more evenly like my toaster? Sounds like a computing machine. Would be nice if something pointed to an organic/chemical neural computing circuitry capable of sentience that we could interface with. Or to the point where we could store are own sentience training an emotion/personality/reasoning engine through osmosis that was a perfect template that could replicate self realization experience uninterrupted? ( forget even singularity… but that sentience/experience storage uploaded to a grown organic host latter sounds like a more realistic long distant intergalactic travel possibility than cryo? )
    Instead immediate trivial business concerns seems to be mandating ( stunting ) our imagination towards market data collection and cheesy robo-servants/slaves.
    ūüôĀ
    Everyday as tech progresses I keep hoping i'll wake up and find out that an ai LOST to the best Russian chess champions!
    And that really upset the ai because it actually cared! That would truly be amazing! ( the endgame? )
    All this fear over ai when in the future they will not be our overlords they will simply be our next iteration. ( and ultimately as we don't live in a vacuum… A force of nature )

    June 27, 2019
    Reply
  5. JJSine said:

    Can we ask AI to successfully design a free energy electronic circuit?

    June 27, 2019
    Reply
  6. Steven Stoffers said:

    so¬†sometime around¬†2040¬†we¬†end up with¬†exactly 5 religions,¬†including¬†sects, or….¬†another 5¬†new ones to add to the pile.¬†many¬†of us will want to know what 'The 5'¬†believe….¬†to get started now,¬†what does¬†Watson believe?¬†is it¬†Buddhist?¬†does it believe or deny¬†Man Made Climate Change? I want to know.

    June 27, 2019
    Reply
  7. logesh j said:

    Will it resolve the poverty issues?

    June 27, 2019
    Reply
  8. LogoMonic Learning said:

    7 people dont understand how important this video was

    June 27, 2019
    Reply
  9. Arthur Dent said:

    Watson might be the first computer that you can talk to while high and get equally competent answers to your input.

    June 27, 2019
    Reply
  10. Ross Gerard said:

    So does that mean the computers can figure out ideas that haven¬īt been spoken by the person speaking just by the way he she talks?

    June 27, 2019
    Reply
  11. Nick Jagodzinski said:

    Interesting and insightful conclusion

    June 27, 2019
    Reply
  12. Sondra Aiken said:

    I do appreciated the investment of money, energy, and marketing that IBM invested into this space. I wonder how they will fare against other juggernauts like Google, Microsoft, and Facebook. Only time will tell.

    June 27, 2019
    Reply
  13. Shadu Shah said:

    why is there a fucking BERMUDA word decorated on the stage.

    June 27, 2019
    Reply
  14. Michal Sas Tymowski said:

    Technology of building chips on neurosynaptic cores is mind blowing even in theory but this revolutionary technology is insignificant compared to quantum computing. Very excited to read about new record of 39 minutes stabe qubit. It is just a matter of time. 

    June 27, 2019
    Reply
  15. Nerian said:

    For languages that are not highly similiar like European languages, google translate sucks for anything more than basic sentences. Just look at Japanese to English, it's equally bad English to Japanese.

    June 27, 2019
    Reply
  16. Graham Hart said:

    I <3 Watson! Watched all the IBM videos about him on YouTube here and couldn't wait to indulge in this TEDx talk by one of his lead developers – Cool beans!

    June 27, 2019
    Reply

Leave a Reply

Your email address will not be published. Required fields are marked *