Neuroscience and Artificial Intelligence Need Each Other | Marvin Chun | TEDxKFAS

[Music] [Applause] [Music] [Applause] [Music] these are really exciting times major breakthroughs are happening in neuroscience and in artificial intelligence benefiting our everyday life in particular neuroscience is making AI better and AI is making neuroscience better and so I will talk about the incredible innovations opportunities and also risks that are happening through these interactions I myself I am a neuroscientist I study the human brain and how it relates to the behavior my major tool for studying the human brain and predicting behavior is MRI magnetic resonance imaging we put participants in brain scanners like you see on the screen and we can measure their brain activity while they're performing different types of tasks MRI has been around for a while as you know it's a very powerful tool to image internal structures in a safe and non-invasive way around 1991 scientists have discovered that you can use these MRI machines to infer neural activity while people are looking at things or thinking about different things and that is known as functional magnetic resonance imaging which I will call fMRI throughout my talk today so I am basically an fMRI scientist and I feel very lucky to be a scientist in this modern age because for us fMRI is kind of like having a telescope or having a microscope without a telescope you cannot study the stars without a microscope you cannot study the microscopic world and thanks to fMRI we can learn more about how the brain operates than ever before you see here on the screen that at least at the beginning in the 1990s Pia we're using fMRI to infer what different parts of the brain do however now in 2018 thanks to AI machine learning machine intelligence and computational power we can learn many more incredible things about what's happening in your brain while you're in the scanner let me show you a few examples this study is not from my lab this is from Jack Allen's lab at UC Berkeley but in inspired many projects in my laboratory this is a study where imagine you are a participant lying down in the scanner you can see a TV screen when you're in the scanner and you're beat you're watching this video on the left okay you're watching a video clip and what what Gallants group did was it they built models using machine intelligence machine learning they built very sophisticated models to understand what is the relationship between when you're watching a movie and the brain activity that you can infer using fMRI and then just based on the fMRI signals alone while you're watching this movie they can guess what you are looking at and that guess is shown on the right of the screen so let me go ahead and run this again while the subject while you're in the scanner you're watching the movie on the left and on the right is a computer guessing what you are looking at it is truly a form of mind-reading okay and this study was published in 2011 and so it's it's a fairly recent you know breakthrough in human neuroscience in fact this technology has been so well developed now that a group in Japan at 80 I have also used this technique to scan people while they're sleeping and while they're dreaming and they have been able to start to decode what people are dreaming as well my own laboratory Alan Cohen and Bryce cool were members of my laboratory they were so inspired by the study that you just saw that they decided that they wanted to try to advance the methods of mind-reading and they did this study where they wanted to see if we can draw what faces people were looking at in the scanner so imagine you're in the scanner you're looking at a face one at a time and we can build models of how the face perception activates your brain and just based on reading out and decoding that brain activity that's collected through the fMRI scanner we can guess on the bottom our computer can draw what faces it thinks that you were looking at when you're in the scanner and although it's a little bit fuzzy you know it for a first attempt it was I think it worked out pretty well the match is about 65% based on what people were looking at in the scanner and what the computer is guessing based on the brain scans again it is a form of mind-reading in a separate project my own laboratory together in collaboration with many fabulous colleagues in particular Todd Constable's group at the Yale University School of Medicine we are using this fMRI method to do other kinds of important discoveries about the human brain and so this is a little different from what I just showed you but let me try to explain what we did here in this study we put people in the scanner and we did not necessarily ask them to do anything in fact we just had them lie down in a in a task known as resting state we just asked people to lie down in the scanner and just rest of course your brain is always active your brain is never idle and importantly everyone's brain is unique and operates in fairly different ways and our big scientific finding our development is that we created a way to convert your brain activity to a spreadsheet of numbers we were able to read out your brain activity during the resting state and convert it to a matrix of numbers a spreadsheet of numbers and importantly that matrix of numbers is unique to each individual it is something that we call the functional connect dumb because what this set of numbers is measuring is how different parts of your brain are communicating with each other in synchrony or out of synchrony but long story short I can take each one of you put you in the scanner scan you for 15 minutes and then I can pull out from your brain a matrix of numbers that will be unique to you and so we're gonna call that a brain fingerprint and that's why for the cover of the journal where this was published they have this big picture of a fingerprint so imagine I can pull out a brain fingerprint from each one of you that's going to be unique to you that alone is not is interesting but what's even more remarkable is that there are components of your brain fingerprint just like there are components of your genome that correlate with different traits and behaviors and so using that discovery we published a finding showing that we can read out your IQ based on your brain fingerprint okay and long story short I can take any one of you put you in a scanner scan you for 15 minutes and then my computer will tell me what your IQ is it's accurate enough that I no longer allow my students to look at my brain activity I don't want them to feel intimidated more over different components of your brain fingerprint correspond to different behaviors for example we want to use this technique to quantify behaviors that are difficult to change into numbers for example in this study we can just by scanning your brain and extracting your brain fingerprint we can read out and quantify how attentive you are we can quantify whether you have attention deficit disorder or now and how severe those symptoms are in other studies we can scan Alzheimer's patients and determine how severe their memory symptoms are we can quantify that we can put an autism patient in the scanner and determine how much cognitive impairment that they are exhibiting in fact in another study our lab can take people put them in the scanner and we can even read out their personality we can determine how extroverted you are or how neurotic you are just by looking at your brain fingerprint measured with fMRI so I'm pretty excited about a lot of these findings and we'll talk about how they can be further used to benefit society but I also want to take a moment to say that this was only possible because of the massive computational power that we started to enjoy over the past ten years and most importantly because of the fancy and very powerful and exciting algorithms that artificial intelligence and machine learning are providing to us as they are providing to the world and we are all already benefitting from these technologies when you're looking for a book or trying to choose a song the product recommendations you get from Amazon or form iTunes and so on are coming from machine learning algorithms just as you might see in this example over here the driving assistant systems or even self-driving cars is possible thanks to advances in machine learning and artificial intelligence and the algorithms for image recognition are becoming so powerful that they even rival board-certified doctors in their ability to classify and diagnose skin cancers or other types of cancers that are measured through radiographs in particular of course everyone in this audience in Korea is very familiar with the fact that artificial intelligence can now beat human champion goal players and and this is really and I'm sure I know it got a lot of hype and attention in the media and what I can share with you is that it has not been overhyped it is truly sensational it is remarkable I think it's one of the most important findings in modern society that a computer is able to beat a human at this remarkable game of Go it is astounding no scientist ever predicted that this could have happened this early in our five times because the game of Go is so complex and because even the most fastest powerful supercomputer cannot explore all the moves possible in order to beat a human champion okay it was only possible because they changed the way these algorithms worked it was because of advances in neuroscience is why a computer was able to beat a human and I would really love to credit demis hassabis the CEO of the most you know this tremendously innovative AI company google deepmind he basically incorporated principles from neuroscience in order to create this massive advance in AI computing and I'm a huge fan of him and it's partly because and I'd like to share with you here is that he was a neuroscientist his training his PhD background was actually an fMRI research very similar to what I did and in fact I I might even say that ten years ago I was more famous than he was but of course things have changed a lot since then he's done a lot more with his life than I have but but anyway his training was in cognitive neuroscience using fMRI and he used his insights his knowledge of neuroscience to bring these tremendous advances to artificial intelligence and I'm very happy and grateful to him because he continues to advocate for this importance of this marriage between neuroscience and AI for example he published his paper where he talks about constantly that his artificial intelligence algorithms are inspired by neuroscience principles and he talks about with so much at stake the need for the field of neuroscience and AI to come together is now more urgent than ever before and he has the goods to be able to say this statement because alphago this remarkable algorithm was able to achieve its revolutionary performance because of this marriage from neuroscience so this is a beautiful image of neuronal cells in the brain and the way to think about the brain is that it's you have all these about a hundred billion neurons they're kind of like little simple computers and they're massively connected through dozens if not hundreds of trillions of connections known as synapses and the most important thing to know about this massive connectivity these networks of neurons is that they communicate with each other and most importantly the connections are modifiable okay and that allows the brain to learn pretty much almost anything it needs to learn in addition the neurons are not just randomly forming some kind of soup in your head but they are organized into hierarchies and they form layers you can think of them as forming layers of different levels of sophistication and processing so if here we have the brain on the top and we have an artificial system on the bottom a neural network an artificial intelligence system on the bottom brain schematic on the top and the idea here is that you can take sensory input vision sounds touch and so on and and you increase when we process that to more sophisticated levels at the beginning you just process sounds and colors and lines and at the end at the other end on the far right side after you go through all these layers you have meaning you have words you have concepts you have faces and scenes just more complex representations and this and this is basically the architecture of how alphago is able to achieve its remarkable level of intelligence in addition there are two principles of neuroscience brain processing human brain processing and animal brain processing there are two principles that are that make us humans so special and they were incorporated by them as Havas into his artificial intelligence algorithms one is replay which is the idea that even when you are not actively looking at something or thinking about something even when you're resting your brain is replaying its experiences and when you sleep when you dream dreaming is actually your brain replaying its experiences from the day and it's not just doing that for fun in the process of replaying its experiences it's forming knowledge and it's forming memories as proof here we have on the top left some pictures of neuronal activity collected from a rat when it's running around in a maze during the day and on the bottom are the same neurons essentially replaying its activity while it's sleeping while it's dreaming okay there's a strong correspondence in fact if you disrupt the dreaming the Rath learning starts to go down it becomes impaired next is prediction your brain you can constantly plan and think about the future and think about the outcomes potential outcomes of different actions and thoughts that you may take and previously computers really didn't do much of you know do much prediction but one of the reasons that alphago can do it's amazing things as well as chess programs is that it has this ability built in this ability to start to predict the outcome of different moves so in sum I think we're neuroscience and AI have kind of a joint mission joint benefit is in providing personalized predictions for humans for machines for any objective that you may have and they're still you know exciting developments are just starting you know it's just they're just starting now I think the next 10 20 years are going to be incredible we're gonna see some amazing innovations that are going to benefit us all and so we need to continue to think about how AI and neuroscience can inform each other number one obviously computers have more speed and power compared to the neurons in your head however humans are much more versatile our brain cognition human brain cognition is much more flexible and it can adapt and it could transfer learning much better than a ughter can you teach a computer one task it's only going to learn how to do it's one task if you teach a human a task it can usually generalize that to many other tasks second computers have this amazing capacity to run simulations like we have to sleep and a dream in order to do replay you can program a computer to replay all the time okay right now alphago was trained on human games the newest version of alphago it's called alpha go zero it can beat the AI algorithm that beat of the champion ISA door there's most current version teaches itself it just runs on its own and plays with itself in order to achieve a remarkable learning and performance that's something that computers can beat humans at you know all the time because they're so fast and powerful whereas with humans one advantage that we have is that we don't need all these training examples we can learn just from one example you teach a two-year-old what a giraffe looks like you don't need to show a hundred or a thousand pictures of giraffes a child can learn and generalize in almost one shot and and that's something that AI algorithms are now starting to learn how do you learn something so quickly in one instance once they start picking that up then AI will become even more smarter than it already is and then the third point is that computer algorithms and AI benefit and can process big data they can process huge amounts of information but they but they don't know what to do with that information they don't know right from wrong they don't know what's good or bad and so ultimately they always we always need a human teacher we always need ethical control we need a human to teach machines what is right or wrong okay because data won't teach a machine what is right or wrong here are some examples this was a chatbot and within a day of release a bunch of human users were interacting with it by Twitter and taught it to be racist and evil this chat bot within a day just because of bad data interactions with humans it started to deny the Holocaust it started to love Hitler and it became very racist because artificial intelligence algorithms they don't know what's right or wrong you know they'll just copy whatever data is being fed to it there's a term for this garbage in garbage out okay they don't have the ability to discern what's good or bad only humans can do that for machine learning algorithms these are pattern recognition systems these are Congress men in the United States and a computer algorithm treated you know categorize them as criminals having consented you know having been in prisoners and of course there are these other very disturbing examples of misclassifications of human faces and also even self-driving cars you know we have to worry about how safe they are because again you know they they don't they're not perfect and you know they need to learn you know how to make judgments when they go down the road same thing for neuroscience we're gaining this ability I can read out your IQ I can read out your personality we ultimately want to use this for good we want to be able to diagnose depression mental disorders reading disabilities there are so many powerful ways to use this especially in medicine for good purposes but we have to be careful not to use this for bad purposes we don't want to be screening people without their knowledge for just certain jobs or certain schools that would be terrible if you got into Yale because your brain scan looked good and you got denied admission to Yale because your brain scan looked bad that would be terrible so we have to make sure we use these new technologies and good ways same thing with AI again AI does not know good from bad and so we have to make sure we continue to have ethical control over our algorithms and in closing again I think the biggest risk with all these neuroscience developments and with artificial intelligence is that is mental privacy okay we should never use these techniques to extract information about individuals without their knowledge if you're going to read out something from your brain if you're going to read out something from their profile on social media or their usage of web traffic then we should be telling people what information we're extracting from them and so I'll close with this I want to thank you so much for your attention and I look forward to the future we have together in neuroscience and AI [Applause]

Channel: TEDx Talks

Views: 24,476

Rating: 4.9452057 out of 5

Keywords: TEDxTalks, English, Science (hard), AI, Big Data, Biology, Brain, Cognitive science, Computers, Consciousness, Data, Ideas, Innovation, Intelligence, Learning, Neuroscience, Open-source, Psychology, Research, Science, Technology, Visualization

Id: 97iYdJE9mQ4

Length: 22min 43sec (1363 seconds)

Published: Mon Oct 29 2018