 Let's start by looking at a landscape of AI. This is a diagram I came up with for when I give talks about learning AI and specifically draw a distinction between what I call research AI and applied AI. Applied AI is things like machine learning, data science statistics, and certain aspects of deep learning. Now, this doesn't mean there isn't also research going on in pretty much all of those topics, but rather that there are practical applied tools that were sometimes called citizen data scientists, for example, can make use of without necessarily getting into the deep theory behind those techniques. Now, when we talk about AI normally or colloquially, what we're actually tending to talk about are these sets of techniques which are most outside. So you have deep learnings with neural networks and so forth. We tend to be broken down further into supervised learning, unsupervised learning, and reinforcement learning, machine learning more broadly. So again, coming up with analysis and results with a large amount of data, typically. And then data science more broadly that a lot of the time involves things like cleaning up dirty data. There's certainly a statistical element in there and so forth. A large variety of techniques in general use a lot of data in order to get actionable insights out of it. And this has obviously been a very fruitful area of artificial intelligence over the last 10 years. Another way to look at that section of the AI landscape is through this lens that's provided by Stuart Russell and Peter Norbig. Russell's at UCAL Berkeley, Norbig's at Google, and they've written one of the more commonly used textbooks in the artificial intelligence field. And they break down the landscape into four boxes. So up the upper left, you have thinking humanly and we'll be spending a fair bit of time in this presentation on this around cognitive modeling, neurophysiology and so forth. Thinking rathnally on the upper right, this is the log-assist tradition. And this is actually where a lot of AI systems like expert rule systems that were very popular for a time really kind of come from this idea that you have a set of rule trees and that you can basically come up with answers through logic. And then the lower left-hand corner, you have acting humanly and this is where things like the Turing test sort of live. This idea that if you're doing the same thing as a human and you can't tell the difference between the computer program and a human being, maybe it doesn't really matter that much how that is actually happening mechanically as long as you are acting as if you are intelligent. And then finally this sort of acting rathnally and this is where Russell and Norvig kind of come down to in their textbook in terms of how they teach AI. This idea of having rational agents that are aiming to achieve the best or the best expected outcome. And they picture this learning agent, this rational learning agent like this. You have an environment, the agent in some way gets the information from the environment through sensors. There's a performance standard, their feedback loops, and then action goes out to the environment in some way. And this might just be a result, an answer or it may, in the case of robotics for example, may actually move a physical part or manipulate the physical world environment in some manner. But again it's basically coming up with the best result that it can. It's being a rational actor. Now we see some echoes of this in other types of fields that have been studied over time. For a long time the field of psychology was dominated by something called behaviorism or behavioral psychology which was trying to answer the question of how do humans and animals think and act. And one of the large figures of behavioral psychology was a Harvard professor by the name of B.F. Skinner who you see in the left here. And he came up with this idea of a Skinner box and basically this box with say a rat inside it. And through the combination of rewards and punishments you could condition, operant condition a rat for example to press a certain button or to otherwise behave in a particular way by conditioning the rat. Now the idea here was that the behaviorists reject the theory of involving mental processes because they basically didn't think that would provide reliable evidence. So it was really just the inputs and the outputs that mattered and not really kind of how you got from the input to the output so long as you did. Now how did this work out? Well if something like the rat pressing for food simple behaviors. Behavioral psychology did a pretty good job of understanding how well not how but predicting what decisions or what actions would be taken but it didn't work as well in more complex reasoning human level reasoning even higher more complex animal related reasoning and behaviors very well. So it worked but there seemed to be a real limit to what it could do. Now as we come back to this sort of whole machine learning deep learning kind of area it's obviously done some pretty amazing things in the last 10-15 years or so. You know for just give a couple of examples again certain image data sets machine learning can actually give better results than humans under some circumstances. In terms of game playing and this predominantly reinforcement learning the game of go was thought to be very difficult for computers and that we wouldn't have a computer go champion like we had a computer chess champion for many many many years and maybe not ever. In fact happened was over a fairly short period of time a go computer was developed it did in fact beat the human go champion and the reason is if you don't know go that well is that go is sort of if you would serve more of a grand strategy game in some ways than chess is and therefore it wasn't as easy to kind of assign values and to understand who was ahead in a particular state of the game. Because you know doing that by hand was often done when programming some of the original chess champions but by pro but by basically showing a go computer a bunch of games and have it play those games out over and over again basically it could teach itself to do to do very well in fact become the world champion. Now having said all that there are some concerns that maybe we're reaching the limits of our current approach to artificial intelligence in general and pretty much all these deep learning techniques are based on some work that was originally gave back to the 1980s and what's happened though is that with enormous computation power with the ability to store vast amounts and move around vast amounts of of data as well as specialized processing like GPUs and Google's TPUs basically application specific types of engines we've basically thrown an enormous amount of brute force at some of these artificial intelligence problems and have in some cases come out with very good answers. However people started to ask you know is this deep learning trick um it is just a one one trick pony what happens if we perhaps come out to the limits of the that technique and Russell and Norvig again uh in their in their textbook uh here's quote from it and we can report steady progress all the way to the top of the tree so this idea that just keeping throwing compute power and maybe doing some incremental advances in neural network techniques might not be enough to solve um to solve more and more difficult problems and you know I think one example you might point to is autonomous driving for example uh make really rapid progress in a relatively short period of time and um you know a lot of people were saying in 2017 or so oh yeah there you're going to have self-driving taxi services everywhere by the end of the next year you know look how far we've come in five years I mean it's ridiculous to think in another one or two that this won't be a solved problem and in fact we're sort of in a trough of disillusionment right now with self-driving and some of the some of the critics of autonomous driving or at least this autonomous driving is just around the corner have turned really to be the correct people now um so let's go back to AI let's pull out our machine learning and deep learning and most of our data science so those statistics for example still certainly play into here and what we have is we have some of the fields they have gone into and really inform AI over time and other fields which are separate disciplines but that very much intersect with AI in various ways um and I'll talk about some of these in a little more detail um I'm not going to spend a lot of time on the left hand side here I'm not going to spend a lot of time on linguistics or natural language processing although this is interesting because this is another one of those areas that um there there's been massive amounts of research as a whole field of study linguistics over time and for instance some of the early voice recognition speech recognition systems really tried to do so from more or less first principles in terms of you know how humans understand speech and what's really happened with our voice assistants and everything is uh everyone's kind of well not everyone but many people just kind of punted on that question for now and say let's throw lots of data at the problem and the results are reasonably good for voice recognition for natural language processing NLP much less so it's uh Alexa tends to understand your commands and I almost shouldn't have said that because there's one in the next room and uh she might wake up but in any case um can sort of transcribe your your speaking reasonably well not as well as a human but really is it you really can't carry on a conversation uh the next two the neurophysiology cognitive science and human machine interactions I'll talk a little more about data privacy uh this is things like differential privacy multi-party computation I have a whole other talk in this uh so I'm not going to go into that here uh and then some of these other things I'll dig into a little bit more uh so you know mathematics is obviously a big part of this because mathematics kind of formalized some of the rules of logic and what rationality is and what expected outcomes are and so forth the the equation here is something called Bay Bay's theorem which is essentially how to make decisions under conditions of uncertainty and update your predictions based on new information and this is going to be very important for some of our decisions some of our cognitive science things that I'll talk about in a little bit um mathematics also kind of gets into questions uh in uh the church touring thesis for example states that a touring machine can compute any computable function um how do we reason with uncertain information so to get statistics plays into this as does probability theory and so forth economics is interesting because I think I mentioned earlier that some of these fields are sort of apart from AI you know though they overlap they have different traditions if you would than AI does so they have tended to be in separate academic departments and the like economics and that's Adam Smith over in the left here you ask questions like how should we make decisions to maximize payouts you know what should we do when others may not go go along and this gets into things like game theory you know what should I do um based on what I think you will do and what you're doing is based on what you think I'll do and this is again a kind of a whole theory of study around when multiple actors are acting in ways that are affected by how other people other actors are doing things how should we do this when the payout is far in the future so the idea that further decisions out or maybe discounted relative to nearby ones economics is also interesting because it's it's evolved from this very rational actor type of approach you know where everybody acts and expected outcomes and everybody is sort of doing the optimum type of thing now I think intuitively humans have known for a long time no people don't always do things they're in their best interests humans have biases and so forth um but there was never really a theory about this and in fact mainstream economists were really rather resistant to this idea that that these weren't just sort of unimportant outliers you know basic economics was rational but then Richard Thaler who I was lucky enough to have a couple of classes with when I was at Cornell um he actually won the Nobel Prize a couple of years for a couple years ago for a surf subfield of economics that's called behavioral economics that really has a theory around utility functions where where people do not necessarily act in a maximize expected value uh rationalist economist sort of way so again so um psychology has sort of made its way into the economics field in fact this sort of behavioral uh different different behaviorism than uh Skinner but this uh behavioral psychology um really kind of infused economics with this sort of new behavioral tradition and finally and I kind of wanted to spend more time in this in this talk but really don't have the time but I'd encourage you looking into this is system engineering and human machine interaction so one of the questions is how do we design in safety at the level of the complete system and you know this is a particular drum of Nancy Levison and others at MIT for example uh in terms of designing safety systems that account for these broader systems including the human in the loop that's an autonomous well sort of autonomous car crash uh over in the left hand side there and when you read about these kind of things in newspapers and so forth you'll often read you know oh the human was at fault because the human didn't do this and this is this is not specific to autonomous vehicles of course it's specific it applies to many many types of major failures and accidents and uh Levison others sort of have a number of tech techniques uh stamp stpa uh which are really intended to draw kind of a broader box around the system include the human perhaps including even the management the culture of company and so forth uh as kind of a way to really you can't just sort of blame oh it's the human messed up it's their fault you really need to look at safety more broadly part of this is out of how do we anticipate the emergent behaviors that are coming out of these complex artifacts controlled by software so um you know again whereas with purely mechanical system you can do this component by component failure analysis but particularly in complex software systems that are instantiated on hardware you can have these complex emergent behaviors that are difficult to predict on a component by component basis and then finally how the human and computer decisions interact in systems with embedded autonomy um this is the this is an you know a whole area of study uh for instance uh at the human autonomy lab at Duke that Missy Cummings who's actually a former uh by the early uh female navy us navy fighter pilot's heads and they're looking at these things like handoff of control you know this idea that you know if you're driving along the highway 65 miles an hour um and you know an autonomous system and works fine 99 percent of the time but you know if the system can't just sort of go oops I don't know what's going on human you have one second to take control and alarms start going off and the like and you know basically that doesn't work and if you pretend it's going to work you're going to fail but what I'm going to focus on the last part of this talk is this neurophysiology and cognitive science area because I think this is an error this is probably one of the areas that we've been studying for a long time haven't made an awful lot of progress in arguably although we have made progress um and in some ways as with as with uh voice recognition for example we've sort of punted to lots of data and lots of computers um because we've been having trouble understanding how humans actually work and this is really the thinking humanly quadrant of Russell and Norvey how do humans themselves actually think now cognitive science as a field dates back 1956 and 1956 is an interesting year because in 1956 John McCarthy up at Dartmouth had a summer workshop that's usually dated to the birth of AI as as a field on John McCarthy who was at Dartmouth at the time would shortly go to MIT and then later have a long career at Stanford Marvin Minsky from MIT was there again one of the other giants of AI um Claude Shannon labs was there um who was very instrumental in a lot of development of information theory so that that was one gathering in 1956 and maybe a couple months later though there was an MIT IEEE symposium at MIT that is usually uh sort of where cognitive science could start the term wasn't actually coined until um almost 20 years later but uh but the ideas of cognitive science were were certainly present and there was three fairly important papers that were presented you know presented at this event and you had a paper by Miller on memory you had Chomsky giving a paper on formal grammars and that's linguistics natural language processing um later uh and then uh Newell and Simon gave a paper in the logic theory machine and this was basically something they had a program that they had written which I believe was presented at the at the earlier Dartmouth uh Dartmouth event and this was a paper and then this often considered to be the first artificial intelligence program that tried to do logical reasoning and cognitive science asked questions like trying to figure out was the cognitive basis for things like learning concepts how do how do how do children how do adults learn how do we decide that things are similar to each other maybe kind of fall in the same category how do we figure out what you know whether a causes b or not how do we form representations from our our senses so we get in this raw data from eyes and ears and touch and so forth and how do we form actual perceptions uh based on those how do we learn word means how does language work um how do we predict the future based on what we know today and how do we develop real world intuitions now um and I'm going to quote Josh Tenenbaum at MIT cognitive scientist at MIT fairly liberally throughout the rest of this presentation and at the very end there's a pointer to a course he co-taught that on that on MIT open courseware that's worth checking out and he sort of said describes as there being two notions of intelligence and the first of these is classifying recognizing predicting data and what we have in the right here is map cholera map by John Snow um that essentially does it these were cholera cases in London and basically by the pattern John Snow predicted that it was a particular well that was the cause of this cholera outbreak rather than the popular theory at the time that there was this miasma in the air and germ theory was just coming in around this time and that's an example of how you can use patterns even if you don't really know the causation model and that that can be useful and this is basically what neural networks do and you know as in John Snow's case humans can do this as well and and obviously for certain things this this uh just pattern recognition is something that works um and it can be useful um so another example that Tenenbaum gives is Kepler's Law Kepler's Laws which sort of describe planets and so forth in the solar system um is really a descriptive model based on Tycho Bray's data predominantly but it doesn't explain why it's like that which Newton's Laws and then later general as layer modified by general relativity does and you know this is again essentially what neural networks do and you've got input layer you've got various hidden layers that do things like edges combination of edges object models and then you get an output now there's still some debate that's been going on for a long time to what degree the human brain actually mimics the way a deep neural network does uh and there's various evidence that yes that you know there there's there's there's an area of the brain perhaps that is associated with face detection and maybe we have edge uh horizontal and vertical edge detectors of various cases in our in our brain um but although there's still some controversy about that there does however seem to be some specificity for certain types of functions in the brain so in the the drawing in the lower left that's showing Broca's area which is uh in the dominant in the dominant frontal lobe is where speech is and Paul Broca announced that uh in France in 1861 so there's been this idea of some specificity of function often detected with for example because of in because of traumatic injuries to the brain for quite a while and in the black and white in the lower right um that's 1957 from Penfield and that that was kind of further trying to identify specific regions of the brain that seem to have certain functions Nancy Camlisher at MIT has been very active in research here using MRI machines to try and try and map in specific functional areas of the brain and in fact there do seem to be fairly specific areas of the brain that are fairly constant from person to person with these kind of speech perception and language and motion and shapes and so forth um interestingly there's no not really evidence in the prefrontal lobe which is where our higher brain functions are for this level of specificity or if there is it's maybe sufficiently abstract that is difficult to detect any differences now Timbom goes on to say though there's also this second notion of intelligence that is really around explaining or understanding or modeling the world that's complementary to pattern recognition um but is more powerful and I think intuitively we can think why that might be the case if we can explain why a causes b causation that is a more powerful concept and say wow a seems to be related to b but it's not clear what the mechanism there is and yeah but it does look like there's this correlation statistically but we don't know why that's not as powerful and this mod or this explanation you really want saying this kind of compact it doesn't vary a lot it's actionable for a lot of things including by by planning we can create plans using that and it's compositional so we can kind of take it apart and extend it and combine it with other things um and ultimately the question that we are kind of trying to answer what you for for just Timbom but the central question is is how do we computers get we humans um really I'm a human not computer I guess so much from so little you know our minds build these rich models of the world and make strong generalizations from really very little input data and that and that data can be sparse it can be noisy it can be ambiguous and it shouldn't be enough for us to come up with answers certainly computers can't do it and you know this is there's a whole bunch of examples and uh again look at link at the end of this uh presentation if you all see some more but you know humans we can look at these pictures so you'll look in the upper right upper left there for example in those bicyclism we can glance in that and give a reasonable count for how make people are probably in that picture a computer would have a lot of trouble doing so and the same is true of in these kind of other other pictures these would be very hard for computers to kind of particularly look at you know some of those people way in the background that we only see part of um yeah how many yes humans we can kind of glance at that and we're not going to get exactly right but we're going to probably get in the ballpark um this is not one that's interesting because humans are you know one of the criticisms is oh people see patterns where they don't exist um and that's certainly true um it's often the case however if you have a causal model for say a cancer cluster or something like that um and that model is in the right you know basically model is how likely is it that given this pattern that there is a causative reason behind the cluster in other words how likely is it to be that this isn't in fact just randomness and it turns out that people are not bad about doing this you know we were looking the upper left and we go yeah there's a couple dots they're kind of close together but there's only three dots and um yeah maybe and you're basically picking a number from zero to ten over how likely it is and you know you then go in the upper right and go yeah that's actually a pretty big cluster there there's probably something going on um and it actually turns out that you run this experiment and humans are coming pretty close to the model um so we actually are not bad and we're just glancing at this stuff and coming up with a number and we're actually doing pretty well and there's lots of other examples of this of uh judging type judging similarity and judging causative causative effects and what cognitive scientists or at least some group of cognitive scientists are trying to do here is come up with what cannon bomb calls a common sense core and this includes things like looking at how children learn for example and the idea of don't approach learning is this thing that machine learning and deep learning models does of data analysis or you know trying to find power but theory building you know a baby a toddler you'll kind of stacks up blocks and sees it falling you know see they fall over and go oh okay that teaches me something about the world um and this idea humans thought structured around basic understanding of physical objects intentional agents and their interactions so physics like the blocks falling over or some intuitive psychology and some of research is pointing to we have some sort of modeling it seems like we have some sort of modeling engine that's built on probabilistic programs but just the probabilistic programs by themselves are probably not a complete explanation of this um there's some critiques of cognitive science i put critiques in quotation marks here because i think many of these are recognized as areas for further study as opposed to something that cognitive science is just deliberately ignoring but you know emotions don't really come in here this is still um within the thinking humanly box it's still kind of a rationalist view to you know how we do things um very importantly it is it disregards the role of physical environments in in human thinking and you know there's there's definitely uh in school of thought that a lot of human learning and development is very related to manipulating the physical world um and similarly it's our disregards uh in general the physical factor in human thought in action there's also a social element um that is certainly addressed in many cases by in psychology but it's probably not a primary thread of social science today and then finally um you know we we can't be you know then we can't just be an organic version of a computer because you know computers fill rooms uh we have this little tenth of a water whatever things sitting on top of our heads so we can't be doing things like computers are doing i mean of course we know we don't do some things very well like multiplying large numbers or factoring primes that computers in fact have very good algorithms in order to do so the brain must be different in some way oh as i said if you all dive deeper this is i was going to put a bunch of links here but actually if you found this interesting i think there's just a really good place to start and there's references to tons of research papers in in uh that course as well so and that's place to go and now we have time for some questions so thank you all very much so if anybody has any questions have a few minutes so Gordon um i have a question a question myself um you mentioned at one point right like the importance of engineering like the systems engineering aspect of it right um how do you sort of see that you know with this whole shift of computation towards like the conical edge um how do you see that sort of playing into it right in aspects of like privacy or machine learning in applied for um engineering these systems do you see that like as a as a possibility what is that song well you know i think the main point well in the terms of privacy as i said encourage you to get look at another talk i gave here uh yesterday at something youtube now but yeah the i think the general i think the general thing and yeah i actually took a stamp workshop with Nancy Levison earlier this summer which was one of the things that got me interested but i think with that and the human factors here act from work i think to a point uh Shreya just made in the chat is serve a um let's not focus too deeply into the it's not all about machine learning it's not all about the deep learning of we really need to take this bigger picture and look at things as complete systems and i think i think in general and you know this isn't even purely the software industry it's you know uh you know what Nancy's examples is uh going through the bulk pile um uh chemical uh del chemical disaster a couple decades ago or whenever it was uh and you know yes the proximate cause was the was a human doing something but there were all kinds of systematic factors going up the stack that really kind of led to this low-level worker doing something wrong and i and you know i think we absolutely see this with autonomous vehicles for example so missy Cummings at Duke has been like very critical of of like any of these autonomous vehicle systems that say you know human needs to take over you know right now and um so so i think it's really just taking a look at the bigger picture um and yeah Williams asking here you know what what are we doing at red hat here um you know i i i think from um you know i think the system engineering part is very relevant to red hat because we you know we have software that runs on complex systems i think we need to think about those kind of issues in terms of the cognitive science part probably not you know probably not so much unless uh we ended up having some sort of connection through red hat research where uh where with red hat research we are doing some work with universities like BU and so forth around things like differential privacy and multi-party computation and so forth so there's various things in those areas that we're touching on with basically with our university programs um does anyone else have any other questions well if you don't thank you all very much for your time and have a good dev comp thank you all yeah thanks for your time Gordon that was a very interesting presentation um just as a reminder um i think we have lunch now i say i think i know we have lunch now so we'll be reconvening at 1 40 p.m edd for our next talk which will be reinforcement learning based dependency resolution by fidel and buccarni William yeah lunch is in your kitchen i'm sure you have something delicious delicious prepared so thank you all and see you after the break