A Brief History of the (Near) Future: How AI and Genomics Will Change What It Means To Be Human

A Brief History of the (Near) Future: How AI and Genomics Will Change What It Means To Be Human

hello everybody I think I'm recording yes I am in a few days I'm going to Los Angeles to give a talk to an audience of oligarchs ie billionaires and allocators of huge amounts of capital billions and trillions of dollars and I'm going to talk about artificial intelligence genomics and how they will affect what it means to be human in the coming decades and as much as I like the oligarchs I thought I would share this talk as well with a broader audience so I'm recording this little video and if I'm doing it right you can see a little picture of me and you can see the four slides that I'm using for this talk since the audience will be non specialists I've tried to keep the figures as simple as possible and what I hope to do is only speak for a relatively short amount of time and then just go to questions and I'm sure there'll be lots of interesting conversation following the main part of the talk now let me start by discussing the current state of AI AI has made tremendous advances in the last five or ten years largely due to an architecture for learning which we call neural nets or deep neural nets which is really quite old so the basic idea that you could create some kind of synthetic network that is modeled after the brain so that it has individual nodes and connections and it feeds information from nodes to other nodes through those connections that idea is extremely old and when I was a young physicist in the 80s and 90s there were already many people working on it but the idea was largely abandoned for these 10 years or so because people couldn't make progress and what has changed though is that we now have much more data available to us we have much more powerful computers and we've dealt a few additional tricks for the training of these networks and as a consequence of those advances we've seen really huge progress in AI to the point where the way I like to characterize it is that almost any narrowly defined task if you can define the task narrowly enough computers can now perform that task better than humans can so among those tasks are things like image recognition so face recognition the ability to read handwriting the ability to recognize people's voices synthesize new voices the ability to read fingerprints the ability to perhaps safely drive a car under you know under fixed conditions or positive conditions those are all narrowly defined tasks which computers are extremely good at and I think we can anticipate that what is meant here by the definition of narrowly defined tasks is going to broaden quite a bit so what people my age have to look forward to in the remainder of our lives is a period of time where computers will be getting smarter and smarter and invading areas of human activity and getting and gradually becoming better than humans at those activities but perhaps not quite reaching what we might call artificial general intelligence or an intelligence which has some kind of common sense or sense of the world a theory of how the world works if you take a little kid and you have them and you show them the world they learn remarkably fast and that speed at which they can learn about the world and form a model for how the world is currently isn't understood from a theoretical viewpoint and there isn't any neural net or deep learning architecture that can reproduce what little kids can do to give me an example if I go to the mall and I walk by a pet store I can say to my son my or a little kid gee how much is that doggie in the window and my son will immediately understand that the dog in the window is for sale that the pet store is a place where people go to buy pets and how much refers to how much money one has to pay the owner of the store to get the puppy no neural net or deep learning architecture or artificial intelligences anywhere near currently being able to understand all those contextual aspects of the world and therefore would have a very tough time interpreting the sentence how much is that doggie in the window so there's a little gap to be closed I think it will eventually be closed experts differ as to how long this will take some people think it's imminent maybe in the next couple of decades or a few decades some people think it might take another 100 years I used to be in the camp that thought it would take on the order of hundreds of years but now I'm because of the rapid progress in a I'm sorry to think it might happen sort of toward the end of my life that would be the most rapid pace at which this kind of thing could happen regardless of what you think the timescale is for the advent of AGI or artificial general intelligence nevertheless humans will be interacting more and more with smart machines those smart machines will have a kind of alien and strange intelligence that is not like our intelligence but again for narrow tasks we will often find it the wise thing to do to defer to the machines so the machine which tells you wow you better go get a check-up because something's funny with your pulse and blood pressure or the machine that says do you better sell that stock because we notice stuff happening in the last few milliseconds in markets all over the world or the Machine that says yes this is in fact a picture of your cousin which we recognized by the similarity to your face and by by comparing it to old photographs that we before in in tasks like that it will become increasingly common that the machine is better than the human and humans will gradually have to learn to defer to the decisions of machines which whose internal processes they really don't understand to illustrate that a little bit into and to explain a bit more what a neural net is I've included this first figure that's on the screen and what it shows is individual nodes connected by different links and the strength of those connections is what is determined by the training of the neural net so you you show the neural net lots of data and you have an algorithm to either strengthen or weaken individual connections those are the lines in the picture depending on how the neural net performs on the tasks that you're training it for and at the end of the day you might have say a million different strengths of connections that have been fixed by the optimization and if you take some simple model like let's suppose the strength of the connection is discrete and it only has ten values you know one is very weak ten is very strong then immediately for a neuron that would say a million individual connections so those are the little lines on the graph you might have 10 to the 1 million different tunings of that neural net and for some tunings it might be very good at recognizing dogs and for other tunings cats and further tunings human faces and each of those without maybe changing the architecture very much each of those tunings will do a very different thing and the reason I'm emphasizing this is because if you actually think about whether a human an ordinary human could understand what the network is doing that boils down to the human actually looking at maybe some list of a million connection strengths the individual strengths through which one of the nodes on this picture influences another and from that figure out something something qualitative about what the neural net is actually doing and you can see right away if we're talking about a million parameter neural net which isn't necessarily the biggest one one could think of one could think of much bigger ones the human brain has many more connections than that you see immediately that there's no simple way or narrative explanation that humans can give for what the machine is actually doing one example of this is the program alphago which plays go better than any human ever has or maybe ever will but even the programmers who built it and trained it don't really know they can't really look into the network and actually say very much about what it's actually doing in detail so this set of technologies that's moving a eye forward very quickly is really phenomenal but it has a character an intrinsic character which is actually more a little bit more like biology than standard programming so if you if you have a program to computer in the past you remember that there are things like do loops and go to commands and all kinds of things that humans can understand what's actually happening as the program progresses along that's quite different from looking at the connection strengths of a million different neurons one is more understandable to humans and the other one is almost impossible for humans to understand so effectively we've come across a technology now that's allowing us to build very complex networks that process information and they can process process information and very useful way we can get to perform tasks that we find extremely useful or important but our ability to really deeply understand what they're doing inside is limited by the intrinsic complexity and so this is just the situation that we're going to have to live with I think it's just going to get worse and worse and eventually we'll have some very strange and alien intelligences around us that are increasingly human-like in a way that they have a kind of general understanding of the world around them and can make kind of general observations or make general decisions about the world not just narrow ones but humans won't really understand what they're doing and so it'll be almost as if some alien species lands on this planet from a different star system and it's obviously they're intelligent it's obviously they can do some things much better than humans can but we don't really don't understand them very well and I think that's the future that we have to look forward to in the next few decades I think it's it's inevitable that we're facing that go on to the next slide the other big trend which i think is going to affect humanity in the next few decades is genomics and because the cost of genotyping has gone down so much in the last decade or so and will continue to go down I believe predictably we're awash in genomic data and we have now fast computers and good algorithms to learn from that genetic data and one of the things that were able to do we're making very fast progress in this area is to predict aspects of the organism from the DNA alone so in principle if you gave me the DNA of an individual human and didn't tell me anything else about that human I could still predict some things about that human whether he was short or tall she male or female bald what color hair what color eyes maybe how smart that person was or maybe even what their personality type is all of these things are already known to be at least somewhat heritable in some cases very heritable and so they should be largely predictable from the genetic information and what this figure shows is the genetic architecture of height so recently my research group used machine learning or AI methods if you will to learn from about five hundred thousand individual genomes and among other things we were able to build a pretty good height predictor so the machine learning train training produce something which can predict the height of a human just from the DNA alone to basically plus or minus an inch or a few centimeters so it can easily tell the difference between someone who is a well above average and high eights and someone who's average in height it's someone who's well below average in height now what this picture shows is the individual locations in the genome where the variants the genetic variants which affect height are located and so if you like it's a map or genetic architecture the human genome and the locations correspond to what you would get if you took all 23 chromosomes and you laid them end to end and the coordinate along that direction would be the individual base pair in the genetic code and the length of that string of chromosomes would be about 3 billion base pairs and among all the different possible genetic variants in that long string about 20,000 are activated by the predictor and each of the individual variants that are activated either slightly if depending on which version variant you have either slightly increases or slightly decreases your height and from adding up all that information the predictor can pretty accurately predict the height of individual people and this has been tested in what's called out-of-sample validation so although the training set was about 500,000 people from the UK biobank we've actually tested the predictor in other populations like Americans that grew up here and maybe many of whom never set foot in England nevertheless it works you know pretty well on them and I give this example a to illustrate that a really complex trait like human height is actually tractable that we can develop methods to predict these traits from genetic data alone but secondly also to just illustrate the complexity so just as in the case of the neural network where we have a neural net that really works well we know it actually does you know properly recognize faces and differentiate between faces of different people the way that it works is you know almost incomprehensible because it's just so complex it is taking pixels from the image of the face and then doing various transformations on them to make features and then combine those features into some kind of classifier which determines whether it's your cousin Al's face or not and so again here what we have is learning statistical learning which results in some very complex predictor you can check very easily that the predictor works by testing it out of sample data but no human will really probably fully understand all the details of all the complex processes involved in this case in determining the height of an individual so it's it's yet another example of machines learning things from data getting producing a predictor or something that's quite useful to us but it will be difficult for humans to understand what is really going on well how can we use predictors of this type or alternatively a related question is how is this ability to predict complex rates can affect the future of humanity and here I give you an example this is an embryo report so this is a report that couples say which is going through in vitro fertilization might order through some advanced genetic testing and the report will predict certain traits associated with in this case embryo for

3 thoughts on “A Brief History of the (Near) Future: How AI and Genomics Will Change What It Means To Be Human”

  1. POULTRY science must be stopped now! Can't you see a 175lbs CHICKEN is stronger than a human being! By 2110 they will be the ULTIMATE APEX on earth and KILL ALL HUMANS!


    Despite my poor English grammar skills and this constant fog in my mind I can't repress myself to howl at the moon the followings :

    1 – This is so interesting, with so much potential and now it's so close and… real/unstoppable! How can the world be focused on something else that selecting embryo for intel… for curing genetic disease and save kids? Think of the children.

    2 – A sincere big thank to the research teams on this field for strongly improving my moral by :
    a) giving a fantastic boost to my faith in the human nature (or maybe the future of human nature)
    b) and an even better boost for my estimations of the probabilities of the human species to survive during the next ("lots of") years.

    3 – Professor Hsu, for a non scientist like me, qualitative and accurate information’s on this amazing subject are sadly rare and your blog and videos are, (considering my comprehension abilities), by very far the best I have found on the subject to improve my understanding on what was coming in genetics.
    And I think I have guessed the end of the present video. So because of you, now, I have a dream that soon smart giant chickens and humans will live together in peace and harmony. Just the idea of it makes my life a great and continuous unfolding, it's a light of hope in the darkness. Thank you for this!

Leave a Reply

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