Social and Emotional Artificial Intelligence
You. Anytime. Waterloo, reminiscing, for the time so, thank you all for coming. It's, really, a pleasure, to be. Hosting, dr. Jesse Howie here at Microsoft Reiser Montreal. So dr. Howie is as, some of you may know an associate professor in the David art Chariton school of computer science at the you know say Waterloo where. He leads the computational, human intelligence laboratory. He. Holds a PhD from, the University British Columbia has, published over a hundred peer-reviewed, scientific papers, and, has. Received. A number of awards including. From, Microsoft have, to talk about that one the. AAI, and, the International Association for pattern recognition, and he's also also, for the American Sociological Association, and I think he may be one of the only computer scientists, that. I know of who has two. Articles, in the American Sociological review which is the best sociological. Review in. States. Possibly. The world so, he's very discipline or disciplinary he's. An area chair for the International joint conference on our initial into artificial intelligence, and was. The program chair and the tenth eai international, conference on pervasive computing, and, he'll be speaking today on social. And emotional artificial, intelligence so please welcome dr. Hawaii. Thanks. Luke, and thanks for inviting. Me here and hosting. Me at this visit so much really is my hometown so I was always nice to come back here and, see. How things have changed a little bit so I'm, gonna yeah. I talked to you about artificial. Intelligence but my perspective. On it might be a little bit different, than kind, of the, mainstream. View of AI these, days but, I'll start with this slide, with some some, pictures of some of the big some. Big achievements, in AI so obviously artificial. Intelligence, has come a long way from. Computers. That could beat, chess, grandmasters. This, is quite. A while ago now - you. Know self-driving, cars and computers that can play much more challenging games, nowadays. But. One, of the things if we look at artificial.
Intelligence, In group. Behavior. Then, the situation looks, a little bit different so this, is a picture. From the, 2017. Robocop. Which is. Probably. The quintessential, sort, of, multi. Agent problem. And. These, robots are you, know are playing on a team and so, it is sort of funny because they're the, robots, in there they're a, little bit difficult to. Maintain. The you know standing. And things like that but they. Do know that they're playing on a team and they probably have some kind of centralized, controller, but, they really have no notion, of what being. A team member means, right. And what. That what they essentially they lack is any kind of idea, of empathy, or culture. Emotion. Or altruism, so they don't have a way of kind of interpreting, what other agents, are how, other agents, are handling, the situation, as. Team. Members as a social, group and there. Has been some, efforts, to make, this happen so let's, try to put some emotional, behaviors. Into our robots so here's, this guy who's he's happy, and now he's not so happy maybe. He'd let in a goal or something like that but of course these are just purely scripted, behaviors, right this robot has no idea what. This means to any other robot and in fact no other robot has any other way of interpreting these signals, okay, and if, you look at humans, when they play games, together you. See this vast array, of emotions going on so this is a short clip from that this, is the the barsa, junior team that just beat this this, Korean, team you can see this incredible. Array of emotions going on and different, cultural, signals. That they're sending, to each other like this guy is using. The face. Slap, which must be some kind of Spanish, yeah. You, know expression, of of, empathy. I'm, not sure it works so well with these Korean guy there he's trying to show. So. Anyways, so, so the bottom, line is that artificial, intelligence has, come a long way and we can do some remarkable, things with it but when we get to the point where AI. Or, robots, or artificial, intelligence systems are trying to interact with humans, or with groups of humans the, situation, looks quite. Different. And. So this is sort of a well-known problem if, you look at self-driving, cars for example, self-driving. Cars have. Some. Difficulty. When they get to busy. Intersections. For example, and one of the reasons that they have difficulty, is that, they can't display. Enough aggression, to get through the intersection right, so this cartoon by the oatmeal is meant, to sort of Lampoon that a little bit and this is a situation where this Google. Car is trying to negotiate passage. With this human through an intersection and, you've probably all been in this situation where you don't, know who's gonna go next and, and somebody. Has got to kind of take, the bull by the horns and kind of drive forward, and that aggression, level is a culturally. Determined thing, so if you go to different cultures people have different ways of doing it right and so, in this situation, the the humans like making, this way too complicated, and the robot has has, no idea of what's going on the, bottom line is that these robots, these, self-driving, cars have no model, of how, a human works okay, so they have no human, model in them they have models of how, things move in the world but they don't have a model, of something like the level of aggression that's, needed to get through an intersection. So. Artificial. Intelligence, has taken. A stab at putting. Emotions, into, artificial. Intelligence, systems and. This. Is sort of coined, this there's a term that's used to describe this is effective, computing so, here I have a picture I don't know why I put Vaughn know I'm in there but anyways, there he is and. He, didn't work on emotion so much but herb. Simon was probably the first person to think, seriously about. What. You know how could you put emotions, into a machine, and you know what what would that mean. And he basically described. Emotions, as being something that disrupts. Cognition. Right so we're, thinking all these high-level thoughts, and we're reasoning, about what to do next and then emotions, come in and sort of kind, of disrupt, everything but. The the the disruption, is there for a reason right, it's Abel's us to better, coordinate with other human beings for example. And. Then, Ross Picard, wrote. The book I guess wrote the book on effective computing 1997. And. The book, starts with this quote. This. Book proposes, that we give computers, the ability to recognize, Express. And in some cases have, emotions. Is this, not absurd right. So in 1997. This, was you, know nobody had really thought this before written. It down and it did seem very absurd, but she was a professor, at MIT so she could get away with saying absurd things I guess and. Since. Then the field has grown and, which is now in I Triple E transactions, on effective computing and there's a couple of major conferences.
So. There's is a fair, bit of work in in the area. And, Ross Picard gave a talk at. Triple. AI in, 2017. Describing. Some of her amazing work on using. Wearables. To detect. Emotions, for. Healthcare, applications. And. They've. Really done some amazing. Things in, in doing that. But. If you look at artificial, intelligence, so for example, in the 2017. Triple A I I looked, through all of the titles when, she was giving this talk and found, the word emotion, appeared three times right. That's, three times in about a thousand, papers and the. Same thing in 2018. And so, I sort, of feel like this this study of emotions, has sort of been pushed a little bit to the side by a lot, of the recent developments, in in a time. But. It'll make a comeback. Okay. So I think that the the view that herb Simon proposed. Which, has sort of been largely what. A lot of effective, computing has, worked. From that. Emotions. Are sort of disruptors. Of cognition, I think, that this may actually be. Fundamentally. Backwards. And I'll. Try to explain why, and then I'm going to in this talk a model, that I think puts the, cart, before the sari, puts the horse before the cart instead, of the cart before the horse and. So. I'll try to motivate it with some. Evidence that we have from neural physiology so. This is a very very highly stylized, picture of what's going on in the human brain and from. This paper by shoe and they guard and what. They do is they describe, two, different systems. That happen, in human, cognition and they, call it a low road and a high road and the. Low road is something that is these are both cognitive, processes, so these are not instinctive. Things these are actually. Things that humans reason about but. The, low road is something that takes, sort, of avoids the large, parts, of the sensory cortex and. Is. Something that happens very very quickly and, is. Able to give human beings, a rough idea of what, to do next, okay. Then. The high road is a. Pathway through the brain that takes longer, but.
Then It goes through this because these cortical reasoning. Aspects. And that's where you do your actual sort of cognitions, and thinking about you, know how the world works and so on and so forth the. Interplay, between these, two systems is really quite, an interesting part of how humans, make. Decisions, and act in the world. With. Again this low road sort of providing, this very fast. Sort, of quick and dirty solutions, to things and this high road coming, in later if necessary. And, kind, of trying to fix things up, right. You. See the same. The. Same idea, of two systems, coming out from. Psychology. And behavioral, economics so perhaps the most famous, example. Of this is this book by Daniel Kahneman, Thinking. Fast and, Slow where, he describes, these two systems again the fast system, one which, operates, automatically. And quickly and then. The slow system two which is effortful, mental activity, in complex calculations, and, the. Way that they work together in, Kahneman's. Were. Or what he's just how he describes, it is that, the system one is the thing it is always in operation, that humans are using to get by and then when that runs into trouble people, call this system too into play in order to help them figure out what to do right, but people do not want to go and use this system to because it takes more energy, yeah. If, you can avoid using it might as well you get by with, much less if you can go with this. And. You, see the same idea throughout, a vast array of different. Disciplines. People running, into the same concept. Of these these two systems, in the brain. Alright. So this, is the sort of what I see is the traditional, view of of. Emotion. In artificial intelligence which, essentially, is what herb Simon was talking about so, you have some of the environment, you perceive, what's going on then, you have your cognitive, systems, that are reasoning, about the, the environment what's happening, so you know you're playing chess and you're figuring out where all the chess pieces are and how you're gonna move them and so on and then.
You That, drives, you to take some action, and then, your emotions, or things that are happening kind of on the side after, the fact so after you sorted, out everything that's going on in the world then, you figure out what emotion you should be feeling and then you sort of have that emotion, and that can modify, how, you're thinking about, the. The task, that you're doing so. Really a cognitive, machine disrupted, by by. Emotion, and again. These disruptions, that the emotion causes may. Be useful right. They may be actually, pointing, you in a direction that, may, be useful to go but. Really they are take. Second place to this this. Sort, of higher level cognitive reasoning, engine. The. Theory that I'm going to describe today really, reverses. These two systems, and I think puts, the affective. Processing, system as the, thing that has the first kick at the can so to speak right so. These perceptions, come in they go straight into our affective, reasoning, engine this, generates, an immediate, action. This is what I should do next okay, and the. Key behind this affective, reasoning engine is that it provides, an action, a course. Of action to take which, is the one that is most socially. Reasonable, in the situation. After. That if. This breaks down then. People, can call in these higher, level cognitive processes. In order to figure out, you. Know how, to fix this right, so. Okay so so, that's the sort of very high level view this, is an emotional machine, which is disrupted. By cognition. Yeah. -. So this this so I guess there's actually, a third system. Here which I'm not putting in which which, kind of bypasses. All of this right so you could, draw another arrow that, is just, immediate, so that's, when if I jump, at you and you jump back that's your your sort of instinctive, muscle memory to avoid. Things coming at you for example that's, not what this is talking about here this is still a cognitive. Process and the. Theory that I'll describe to you is this. The, this. Cognitive process relies, heavily on. On, language. So. On how we interpret, the world in terms of of, symbols. Essentially. Right or you know linguistic. Concepts. Okay. Any. Other thoughts. Or questions. Okay. So I'm, gonna now describe, to you this theory and then I'll give you a few examples of how I've, gotten. It working. In a couple of different contexts, so. How we're gonna do this is, I'm gonna look back at this work by, Osgood. In the 50s, so, Charles, Osgood was a psychologist, who did. These amazing, experiments. And he went around the world and asked. People to rate, concepts. So basically words in the language but he did this in many different cultures so different, languages, and, he. Got people to rate them on what. Are called semantic, differential, scales, and semantic. Differential scales are just a scale with two opposing. Words at each end right, so, angular, 2 round it or week two straw in a rough to smooth right. And he. Asked people he'd give them a word like in this case the word polite, and say, put a check mark on each of these scales where. You think politeness, sits, okay. And what's.
Amazing Is that there's this incredible. Degree of consistency. Amongst. Humans, that share the same language that are in the same cultural, group as to, where these, words should go so. This picture shows something, one of the diagrams. From his study in 1952, showing. Two groups of 20 people this is the mean of two groups of 20 people and you can see this amazing degree of consistency, everybody, agrees, right. So. What's amazing about that is that if you look at the word polite, and you look at some of these scales so you might wonder well. What does politeness have to do with being rough or being smooth, right, like it seems to be unrelated. In some way at, least on a sort of semantic, level but. Everybody. Agrees yeah politeness, is more sort of smooth than rough right. So. There's something really interesting going on here that everybody, has this, implicit. That we all share, these. Ratings. Of things on these different. Scales. So. What he did was they did these studies and he actually used many more dimension, than this so there was I think on the order of like a hundred and fifty dimensions, that he used so 150, of these scales. And then, he went back and did a dimensionality. Reduction and, found that he could account for over 80% of the variance, and all this data with, just three, dimensions, okay. And the. Three dimensions, are one dimension, of evaluations. Good, to bad, one. Dimension, of power so, strong too weak and one. Dimension of activity, so sort of hyper two asleep and I, I'm gonna put this one coming out of the board so bigger things are more hyper. And. So. When you get people to rate words they, all agree, on where they should go in this three-dimensional space so. For example being polite is good and kind of powerful and being, abusive is sort, of weak and bad and really active, right so this is a bigger word. And. Being serene is good and kind of weak and really inactive. And. You can do the same thing with identities. So. Students. And professors, for example, is a good there's a good example so students. Or everybody. Kind of agrees that, students, are our. Students. And professors are about the same goodness, right, but maybe professors. Are a little bit more powerful than students, and students. Are maybe a bit more active than professors, right, and. This, is culturally, dependent right so if you go to Japan.
For Instance then, professors. Are typically, seen as significantly. More powerful, than students, right. So but, if you come to North America then this has sort of become normalized, students, and professors are seen as much more equal. On this power to mention and. Behaviors. As well so apologizing. To somebody and fighting with somebody and so, on so this is actually real data taken, from a study done, in 2003. In the United States so these are the, means of put. The words that the means of where people, would rate these concepts, again. I'll just reiterate that the important, thing to remember here is that everybody, in the culturally, in the cultural group agrees. On where these words should go in this three-dimensional space. Okay. So how, we're gonna take this this, idea and leverage. It and turn it into a control, principle, that we can use to drive an artificial, agent and we're. Going to do that using this a fact control theory, and what. Effect control theory says is that people carry around these fundamental. Sentiments. Okay, the fundamental, sentiments, are these things here this is everybody's, agreement, on where things should go in this, very simple, just three-dimensional, space. When. A social. Action occurs, so something, happens, in, the world so, for example maybe in this case here we have a nurse and here's a nurse for dumping. Some water on this guy's head and here's, a nurse helping, this this lady for example these. Social actions create, what are called transient. Impressions. So the transient, impressions, are also, culturally. Shared things they, are how, you feel, about the situation so, how you feel about these things in context. All, right and the. Transient impressions, may be different, from these fundamental sentiments. And the difference, is called, the deflection, in this theory and, the. Effect control principle says that people try to experience. Events that confirm, fundamental, sentiments, so, essentially they try to reduce, this deflection, so. Effect control theory is a consistency. Theory it says. That we will try people will basically do, the thing that. Sort. Of leads. Them to situations, where things are emotionally, consistent.
And The, consistency, is defined, by. A. Culturally. Shared, construct. Which I'll show you in a sec. Yeah. So any okay so I'm going to go through a couple of levels of sort, of increasing. Detail of this this, theory and this is sort of at the the highest the, highest possible level, but. Interrupt me any time if you have questions as we go but I'll show you a bit, more detail of this as we go along it's. All described in this really nice book by David hi so who's the I guess. The guy who's done. Many many years of work on this theory starting. From the work of always good and many others as well and. Described. In this really nice little book called expressive order one. Of the interesting things about this book is that it comes it sort of has two parts the, first part is all the social psychological part, and it's. Just you know it's text and it describes his work in in looking, at a variety of different cultural groups and and how, they follow these kinds of dynamics, and then the second part is as a mathematical. Model okay. And what's, sort of interesting is when you talk to most sociologists. They've sort of all read the first part of the book but come kind, of ignored the second part so they you, know they don't tend to be. Math. Amend. To have that sort of math background, and. When I read it I read the first part of the book and that was like completely lost and no idea was going on but then I read the second part I was like oh this makes total sense right, so, this. Is why it's good to have a multidisciplinary. Group. Working together. Alright. So here's a little bit more detail affect control theory uses, a very simple grammar of an actor behaving, on an object, the, object is often, another another. Person, right. People. Have these fundamental, sentiments, about each of these three things right. So this is a nine dimensional, vector and it's, according. To the theory these vectors, lie in the range minus four, point three to, plus four point three and I've, asked every sociologist, I've met who works in this theory of why it's four point three and. Nobody seems to know I think it's lost in the sands of time but, it does have something to do with Osgood's, original, experiment it's actually not important, and the versions that I work with is not restricted. To be in this range and you could rescale, it to be anything you wanted. Okay. When, you see an actor behaving on an object, it creates, a transient, impression, for each of them as well and this, is again gives. You a nine dimensional, vector in the, same space. The. Predictions. Of the transients, from the fundamentals. Are, given, by this nonlinear. Operator, which I'll show you in a second and, this. Is something that is also culturally. Shared everybody. Agrees on how these if these impressions, are created, the, deflection, is then just the Euclidean, distance between fundamentals. And transients and the, effect control people principle, says that actors. Work to reduce. This deflection. Alright. So you can use this as a control principle because you can say part. Of this, these. Sentiments, is a behavior, so, you can say what. Is the best behavior, to take in, order to make, this actor, and this object, agrees, sort, of emotionally, to be doing something emotionally that's consistent, okay, so you can use this deflection. Measure, as an optimization service, that you can walk down to, find. The optimal behavior to take. Effect. Control theory also has a very precise definition. Of a motion which is remarkable, this is I think the only theory I've ever seen it has such a precise definition the. Emotions, are described, as being proportional to, the vector difference, between, fundamentals. And transients, and people. Use emotions, as a, communicative. Signal. To. Tell the, person that they're interacting with where. That person is getting it wrong right, so. If I'm interacting with you and you're, not treating, me how I think I should be treated in this situation, I'll send you an emotional signal, in order to help you correct, for that and so, that we can come to an agreement on, how we're supposed to be behaving yeah. It's, a vector in the same three-dimensional, space. Are. These points. Yeah. There would be a point in the space yeah and if you look back at here. That's why these emotion, words are in the same space right. So. The nice thing about this three-dimensional, space is that you can compare, all these different, entities in the same space so behaviors.
Identities, Emotions, they're, all comparable. Directly. Right, with a simple, three, dimensional vector. Okay. All. Right so here's the. Little. Bit more detail, so we have a social, interaction happening. Here here's our nurse and our patient, we've, already done a population. Survey, so we've asked, people, what. Do you feel about a nurse okay, and we've gathered, a rating for nurses, which. Is this one here so nurses are seen as two, point nine one point five zero point two for example and. Then, we've asked them about patients, and patients, are a little bit less good much, less powerful and much less active, and how. About comforting, the behavior of comforting, well that's positive, and powerful and not very active right. Given. These fundamental, sentiments, we compute, a. Transient. Impression, for each of the three and the transient impression, is given by a series, of. Equations. With. Polynomial. Features. In them so the polynomials, are basically these polynomial, features are combinations. Of these different, elements. In, the fundamental, sentiment, okay so this one here is the evaluation. Of the behavior, multiplied. By the evaluation, of the object, and then, it has a positive, coefficient and, that's, the effect on the evaluation. Of the actor, right. So, what this term says is that if, the. Actor. Does something, good to, a good person then, they will seem more good, right. These. Coefficients. Are things that are also culturally. Shared and measurable so we can ask people how. Do you feel about a nurse, how do you feel about a patient, how do you feel about the act of comforting, how. Do you feel about a nurse when you see a nurse comforting, a patient, okay. And that. Measures. This transient, impression, and then they can do a regression, and find these coefficients, and find. That they are again. Shared, culturally. All. Right so the difference, between these two things is the deflection and like I said before I can use that to. Predict the next behavior, so I can optimize over this deflection, measure in order to compute, the best behavior to take I can, also compute, a reaiiy denta fication of the, nurse and the patient so. I can say if I, see somebody comforting, a patient who might that be Wow, well that's probably a nurse right. So. I can do that as well okay. Before. We go to an example which might help clarify some of this any. Questions. All. Right so let's walk through an example we'll, use this nurse patient example, here. Is our nurse and our patient, again. Here is their 3, D evaluation. Potency, and activity. Fundamental. Sentiments, and remember. These are, culturally. Shared I'll just keep reiterating that so we're sure we remember. It the. Nurse is up here and, the. Patient, is here the important thing is that the, nurse knows. That she is here and knows, that the patient is here and the patient knows that she's here knows that the nurse is here ok, they both agree on that. So. I can compute well let's see what happens if the nurse comforts, the patient, right, so I compute, the transient, impressions, which are in light blue and you, can see they're quite close to the fundamental, sentiments, and then. I can compute what emotions, would be felt. And communicated. The, nurse who comforts the patient would feel compassionate. And the patient would feel serene, okay, and the. Deflection is low is a 1.3 so that's a low deflection. But. Let's say we make the nurse ignore, the patient right, so here's a nurse who's ignoring, a patient ignoring, is is quite, negative, and not. Very powerful. Now. The nurse. Seems, much more bad, than, a normal, nurse would seem right, we'd, say well who's this this is not a very nice, she's ignoring this poor patient right but. The patient also seems more bad because, well maybe this patients been a really nasty, patient. Right and that's why the nurse is ignoring her so we construct, all of these things but, we do so just by, computing, these transient, impressions, it's a very simple calculation.
The. Emotions, that would be created, or agitation. For the nurse and indignant, the patient, feels indignant, how dare you ignore me I'm the patient right. This. These are the remember these are the signals that one that, each would send to the other in order to correct, for this so. The patient, would send this. Vector, to the nurse and say. To the nurse essentially, by sending, that emotional, signal which humans can do very very quickly. Hey. Listen I'm supposed to be up here and you're treating me like I'm down here so. And I'm telling you exactly how you're treating me by sending you this vector and the, nurse can take that and say oh right. I've got it wrong I thought you were you. Know horrible patient, but really you're not right. So I'll correct, for that and do, something to make it better. Deflections. Hi. Yeah, so, okay. What it means on the county on, the behavior, is that you. May be interpreted, at the behaviors, right so if your team is happening, then, you're like me. I'm, sure that's a nurse and I'm sure that's a patient maybe she's not ignoring the patient right, people. Are you seek a an. Interpretation. Of the situation that makes, it emotionally, consistent. Right. So the transient impression, would be one way of getting. That from the from the behavior right. So. For example this this is gone what, you can see here is this is going up in. Evaluation, right, and it's. Going up in power so, one possible, so that's saying well maybe this is not I don't know what word would be close to this transient. Impression, but it's probably something less. Like ignoring, it's nothing more positive, right, well maybe she's busy and just will, come back to this patient later or something like that. That. Make sense. Alright. Here's. Another example this is a student and a automated. Tutor so this is one of the early examples that, we built with this theory and the. Student, and the tutor are, this, is kind of like student and professor right they're kind of in the same spot but maybe, the tutor has a bit of power over the over. The student so. We can compute well what would a student do to to. A professor, well the student would do, something, like this. EP and a rating and I can look up in my, dictionary. Of sentiments, that I've had rated by people and find this, behavior, well that student might ask something, to the tutor. And, then. The deflection, is low the transients, are close to the fundamentals, and everything's good. What. If this the tutor, yells at the student okay, this, creates, transient, impressions, that are very far from these fundamentals, so this well, maybe this student, is really you. Know a lot more less a bad student or no but there's a horrible tutor right, and. Then the the the obvious of what emotion would be felt well maybe this tutor is angry and this student would be exasperated. Or remorseful. Then. We can compute well what would the student do in order to try to correct for this situation, well, they would maybe listen, to the to, the tutor, and feel, self-conscious, and this, act by, the student, would help it to restore his. Fundamentals.
Transient. Impressions back towards his fundamental sentiments but not so much for the tutor because the jury's still out the student doesn't know maybe. This tutor is really a nasty tutor, after all, but. If the tutor then does something so this is the the computation. The optimization, for the deflection so it's minimizing. This deflection well. What would the tutor do something. Like apologize, to the student and now, they've sort of managed to get back to somewhere where, the deflection, is fairly low. Okay. So what I did here in this example was I forced. The tutor to yell at the student that's not an action that effect, control theory would compute because it creates. So much deflection right, so, I create. That artificial. Action, and, and compute, what's the optimal, thing for the student to do to try to fix, things and then, what's the optimal thing for the tutor to do to try to fix things even more. Yeah. It could there could be circumstances where it's appropriate yeah, and people would could, yell at each other for a variety. Of different reasons one of them is maybe, not yelling at maybe, that's not that's too a little bit too strong yeah. One. Of the things that happens when deflection, goes up is that, this system too thinking, kicks in right so. When there is a inconsistency. On the emotional so if the basic idea is that if I'm operating, this, is sort of like an equilibrium, theory, right, so, if we're on the same page and we're, operating according, to this emotional, model as long as everything stays at, that, equilibrium point then. We'll just keep doing that right it's really easy you don't have to think you just have to operate in this simple three dimensional space as. Soon as things go off the rails though then, you kick, in this higher level cognitive reasoning, right, and one, of the things that, that. That might help with is that it might force the student to actually think about what they're learning for, example right, and then, they'd actually consider, it and take, the information in instead. Of sort of just ignoring it and pretending, like everything's, fine. So. That could be one reason yeah. You. Just said does. That sort, of is, there, an implication maybe that you. Some shared the scenarios. It. Has there's, utility, and having high deflection, because it forces you to engage. The. Higher level system and start thinking. Yes. No. No there could be for sure yeah and I'll, talk. About that bit more in a second but effect control theory. Doesn't. Have, a utility, function right. The. Version that I've developed which. I'll show you next does. But. A fat control theory is only describing what's happening at that equilibrium point right, so let's talk about utility, again after I show you Bayes act. Right. Right. Yeah. Yeah. That's that's but a student staring at the window is not being a very good student right so they're already off, the off the mat off the you. Know the equal that equilibrium point right. Is. There a circumstance, where. Yes. I'm you know I try, to think of a good. Example there, but. Yeah. We're. Just looking at yells at and I'll bring to you like. Lower. Left quadrant, so, the student was in like the upper right yeah. The. This this like. Going getting. A transient that is here from here is only based on this point so if the student was actually up here then yelling at wouldn't necessarily bring them in the same spot right so, so but you're right it could be something along those lines right maybe a better example is. The. The work that's done in. Cognitive. Therapies, for depression, so, a depressed, person is someone. Who has. A very negative. Sense. Of who they are right so they feel kind of negatively, about themselves, and for. Many years people, thought that the. Way to handle this was to blast the person with positive. Positive. Thoughts right. And, that, doesn't work very well and the reason it doesn't work very well is because you're not treating, the depressed person, like a depressed, person that's, how they think they are right. And so, you have to so, the self verification, theory came along and said that well what you need to do is agree with the person that the world is terrible and that everything sucks and and then. You say yes this is the way you are I'm I am verifying.
Your, Your. Sense. Of who you are as a depressed person right so I'm performing negative, actions, but, I'm doing it in a way in order to get us basically at that point of. Emotional, alignment. And then, from there I'm gonna slowly try to work you in a direction where you, hopefully. Become less depressed, for example. That. Make sense. Okay. All. Right so my work in in, this, has been to develop so, a version that adds, a, couple of new pieces to this one is a utility, function and the, other one is modeling, identities, and behaviors as probability, distributions, right, so, of course it's not the case that everybody, does agree, exactly on, where these concepts, sit in this three-dimensional space so. When you make these measurements you actually get a fairly, big variance, around the mean so, everybody does agree on the mean but, there's a lot of disagreement on, where they would go okay and this, various, turns out to, be very very important, so, it's something that people I believe use. In order, to handle. Operation. Operating, in groups right. So the problem with group behavior, is that, everybody's. Different there's this diversity. Right we aren't even if you're in a single. Cultural, group so we as academics are, often in groups that involve many different cultures, but. Even if we're in a very uniform, cultural. Group there's, all sorts of differences because people are interacting with each other in many different circles, and it's sort of maybe in a small world type type, social. Network, because. Of that there's, lots of uncertainty, that comes into play and we need to be able to model that uncertainty in order to, in. Order to use this kind of, idea. This emotional, consistency, to cooperate. With other people. Okay. So those. Are the two pieces and I'll walk you through how this works as well I have to have. Quite a few number of papers on it but two, of the journal papers one in artificial intelligence the other one as Luke pointed out in the american sociological review, in. This paper won two awards at the aasa', meetings in in, 2016. All. Right so the model is based on a partially observable Markov decision process, or give. You how what this is all about and then we'll see how we can fit this effect control theory piece on to a generic. Model of how an, agent interacting, with the world so a, pom. DP or partially observable Markov decision process, consists, of a set of states X, and, a. Set of observations, that we can make about those states so these are unobservable, or, latent states a set, of observations we. Have an observation function, which is the probability distribution relating. These two things there. Is a set of actions that an agent can take that. Change the world according to some transition, function which is another probability. Distribution, and then, we have some, reward. Or utility, function, that, describes, the, agent's preferences, over these states. Using. A numeric scale, okay. So that's. The generic model. And. You can use this view a number of different things and. That. May be the simplest, way of visualizing, how you can use a pom DP to, generate. A plan is to, just do a very very simple forward search right, general this doesn't work but this is a way of describing what you actually are trying to do so. Here we are in our current state and remember. This state since. This X. Our. Sort of interpretation. Of the world is latent. We can't observe it directly then, the state that we're in is actually a belief state so this is a probability, distribution over all possible, states, from. This state my, agent, has a number of different actions that it can take which. Lead it to another. Belief. State after, taking the action and then.
The Other person that the agents interacting, with has, a bunch of things that it can do and that leads to another state and so on and so forth and. Down. Somewhere, in this tree is some rewards, let's say and then I can draw, this whole tree, out, find. These rewards, and then propagate, them back up through the tree weighting them by the probability, of these things happening, and then, I can use that to rank the actions, that my agent has and pick, the one that optimizes. It over the utility, function. So. That's just basic Bayesian, decision. All, right so what's the problem with doing this of course is this tree you can't write it out it's way too big and these branching, factors are colossal right, and so, if you tried to do this you'd rapidly, run into, major. Problems. And a. Lot of the work in reinforcement. Learning and decision theory is how do you expand. This tree in a smart, way so, that you don't wind up having to to. Draw. The whole thing out so. For example you can do things like say, well, as, far as what the other person is gonna do I'm just going to ignore all the things that are really unlikely okay. That could reduce the branching factor here by doing that but. This branching, factor is much harder to handle because I can't ignore any action, that I have available to me and the. Reason is if I ignore one action, it could be that that is the action that leads to an infinitely higher reward, than anything I've seen before. Right. And so, that's the reinforcement, running problem I've got to try them all if. I don't I might miss the one that is gonna be you know. Utopia. All. Right so now we're gonna put in this social. And emotional reasoning. Engine and see, how we can use that to help us guide. That, action, selection right, remember back to the start of the talk we talked about this. Emotional, reasoning engine is gonna sort of tell, us roughly what to do in order to get by right. And we're gonna do, a bunch of maybe cognitive, calculations, after that too to, help us out so, I'm gonna take my basic, palm, DP and I'm gonna just make a slight modification in, order to make the presentation a little bit easier I'm, gonna add a random, variable, that describes, my own action, all right I'll call this B Prime and it's, just deterministically. Set by a okay. So this is no, change at all in the model I'm just adding this this, one random variable and, then. I'm gonna layer in a set, of random variables one for fundamentals. And one for transience, and I'm, going to connect them through time okay. So the transients, are going to evolve through time according. To this impression, formation equation. So, let's say well what do I expect. The transient. Impressions, to be after, I see an event taking place the. Fundamental, sentiments, will evolve through time as well I will admit the fact that these fundamentals. May be changing, slightly so. If I for, example go, and start living in a different culture then, I need to sort. Of shift to my fundamental sentiments. A little bit so I'll admit that maybe happening. Then. I have this red link between the two which, is my effect control principle, it says that I.
Believe. That fundamentals. And transients, will be close, together and. Then. I have these green links which, are my, shared, cultural, sentiments, about things so. Essentially, saying that things. In the world have fundamental. Sentiments, okay. And that's another probability. Distribution, and that can take into account the variance in these fundamental, sentiments. So. Given all that I can combine them together and compute a probability distribution, over, fundamentals. Given. All of these things and then, I can, integrate. Out the. Identities. And create. A distribution, just over, the, behaviors. In this three dimensional space, okay. This. Probability. Distribution, I've, actually labeled, as pi because it's a policy. In a way it tells me what. To do next this is a probability, distribution in, ep. And a space but. It's telling me this, is the emotional. Signature. Of the thing to do next in this situation, right. As a probability, distribution so. Then I can draw a sample from this and that's, the thing I should do next. Right. Or or I can seek the maximum or something like that. That. Makes sense. Okay. All. Right so how does this help us well, essentially from a belief, state it tells. Us this, function, here says. Well, you know what you don't have to actually consider every, possible, action you. Should consider actions, only that have this. This. Emotional. Signature, right. Because, those are the ones that make sense to, the people that you're interacting with there's, a socially, acceptable, behavior, that, you should be taking. So. The way I imagine, this is that imagine, this huge tree is in the dark which it is right because we can't we can't draw the whole thing out and this. Effect, of reasoning, engine is like shining a beam of light down. Through this tree okay, so it's saying these, are the things you should consider next. Don't. Even think about these things over here right. So it's an intentional mechanism. It's. Not a way of ranking all of the actions, it's. A way of saying if, I draw a sample from this distribution I'm, going to get something that's in this flashlight beam and the, things in this flashlight beam were the things that are going to work at this, emotional. Consistency, level. Okay. So. Essentially this is a way of solving this problem by. Having. This shared cultural, information on the side that we're using in order to guide to. Guide attention. Okay. This. This same idea actually, comes you. See it in another. Branch. Of sociology. Called practice, theory, which. Is sort. Of it's, fairly far removed it seems from effect. Control theory or at least the practitioners, of these theories would would, say so but. What's remarkable is that board, use may be one of the big sort. Of originators, or thinkers. In practice, theory. Described. Something. Called the habitus, which is very, very similar to what, this is doing so, I'll just read you this quote responses. Of the habitus, may be accompanied, by a, strategic. Calculation. But these calculations are first defined, without any calculation in, relation. To objective, potentialities. Immediately, ascribed, in the present, things, to do or not to do things, to say or not to say okay. So, those immediate. Those. Objective, potentialities. Or, we. Can describe them mathematically. Using this effect control theory okay, and what. He's describing this strategic, calculation, is that this flashlight, beam is, not just. Illuminating, a single action it may be illuminating, just a couple of actions right. And within. The flashlight, beam I can, then use my cognitive, reasoning processes. In order to try to find what. Is the best actual. Action, to take. In order to optimize, over this utility, function. Right. So. This we get back to the utility function that you were asking about earlier. I'm. Still gonna try and optimize this right, but, I'm only gonna do so within this subset. Of these branches, of the tree. All. Right so the idea is that the humans will only do this if they have time, right. Time is an important thing if time runs out I've got to act then.
I'll Just do the thing that's most emotionally, consistent, if. I have time to sort of think about a little bit then, I'll wind up maybe finding, something that benefits me a little bit more, right. People. That that do too much of this tend, to sort. Of throw things into disarray in social situations but. That's. It. Goes, very much in line with what Buddha was talking about I believe. Okay. So. That's a that's all that's sort of the, model presentation, that I'll give and, I'll walk, through a couple of examples actually I'm already at almost, at a time so, I. Might. Just go through the first one and then briefly the second one but any. Questions about the model. Yeah, just, going back to the original. Sort, of three-dimensional. Embeddings, let's say of words. They see how. Many how. Much coverage, of. Like. The Lex, there. Like, how many words have one, of these associated, three. So, there's, a number of different basically depends on how much you money, you want to spend on your study right. But, the, sociologists, have gathered large dictionaries. So the. Most recent one that was gathered has I think. 800. Or somewhere, somewhere. Around 800, identity. Words and 800. Behavior. Words and. They're. Rated, by on the order of 50. Individuals. Each one so you have a pretty good sample of how, people feel about it. Yeah. So, those those dictionaries, exist, there's also bigger emotional dictionaries. That. Are on the order of maybe or 10,000, words or something like that but not gathered by these same after control theorists so. Embeddings, - what. We use more regularly, with word. Emotional. Language modeling. Yeah. Yeah. So we actually so I've had a few students work on that problem and I know a few people who are working on it so exactly that that ideas that would take a take a word embeddings, model. And somehow also integrate, this it, may be the case that if I construct, a let's. Say a three hundred dimensional word, embedding that these three dimensions, of EPMA are in there somewhere right and and, so. It's. Not altogether, clear that that is happening. Though, because you you. Wind. Up with like. The movie. Board. And betting strained on the movie, ratings for instance you get the, word good and bad or really close together right, because they are used in very similar contests. Yeah. That's right yeah so I'm not I'm not clear, on on, that yet but.
Right. Yeah. Party. Blanking the terminology, you using the kind of the initial emotional effects of better than this two dimensional space are are. Are not dependent. On these. Relationships, right there's just a word for two there that's right. So. Have. You thought about. How. To. Do. This or think, about this in. The context, of those changing relationships. Yeah. So I mean that, is. A theory. So. By. The. Concept, is that I want, to say I woke up in the hospital and, I see a guy you, know a person, in a white lab. Coat right I will, have, a sort. Of my, cognitive. Process will say oh doctor. Right this is a doctor, and then, I'll associate, an emotional. Fundamental. Sentiment, with doctors, right this is how I believe doctors, are and then. Using that I'll compute what I think this doctor will probably do to me right. But. That concept, that fundamental, sentiment, is going to be the same in any. Context. Right, so if I see a doctor when I'm taking the subway then. I'll be it'll still be doctor, but, then it'll be doctor, on the subway right, and so, that additional context. I can put in as well and that, context, would change things further. So. Let. Me reiterate, that when I described. A CT. Remember. We just had this three dimensional. This. Three dimensional grammar, of actors, behaving, on objects, so I can augment, that and then this has been done with a fourth. Thing which is the setting or the context. Right so, I can say actor. Behaving, on object, in some setting, right. So student. Joking. With professor, in the. Classroom, is different, than student, joking with professor, in the bar right. Because. It because, it's changing. Over time I mean just. Listening, a single, clanging you can bet the entire state within this context, but, that seems much more difficult, to do then, you, just have the, relationship. The, you know the emotional content be able to change over time. Yeah. So the, tricky thing is that the the the state. Of the, MVP right, this underlying, denotative, state this X becomes. Is very can, be very high dimensional right this this element, here can get really complicated, okay. But, the key thing is that these. Pieces. Here, always, stay the same they're, always in this three-dimensional space and they're comparable, within that three-dimensional space right. So I don't have to, model. All of this context, if I can just get this emotional. Meaning. Of things so. I can compute a whole bunch of things just with the emotional meaning without, having to figure out how the world works right. Okay. So. I think I'm out of time, so I sort of I think it gave this much more slowly than I usually do but that's -. What do you want me to do should i. Okay. Okay. I'll skip this prisoner's dilemma stuff, but, this. Is sort of a canonical of. The sort of minimal working example, of using this theory in practice and I can actually show you a demo of this game. Play. If you'd like after. The fact and I'll. Describe. The. The other piece. That I wanted to mention was the famous COG, project, there's, a project that's in collaboration, with my. Calling to be a Schroeder who was the guy who introduced, me to a fat control theory and he's, in Germany at. The. Potsdam. University, of Applied Sciences Kim. Rogers who's a sociology, at Dartmouth College and main. A gapan who's a software, engineer. With. Me so this is funded by this transatlantic. Platform. Digging in a data challenge, and what. We look at here is we're looking at. Interactions. Of people on, online. Collaborative. Networks. Like github for example so that's our primary sort of target so, we look at people. Interacting. Over you know pieces of code on github and then we're really interested to find out how, do these social, and emotional factors, come into play when, people are trying to actually do something together in. A social network. So. I won't give too many details about, that so maybe I'll just skip these last, bits. And. This. Is or this is the effect control theory version of the, group simulator, so I can create a instead. Of just doing an interaction, between two, agents I can do an interaction, between a whole group of agents and then, the blue agent, is the sort of everybody's. Concept of the group as a whole and I can use effect control theory to compute, not, only what people will do to one another but who they will do it to all, right and the basic idea is that people will try to interact with other people that, are going to help them sort. Of confirm, their sense of self the most. So. We can do some pretty interesting group simulations, with this. Then. We've gathered some data sets as well so a few, data sets of that we have a new data set of. These ratings, of actually. Github, concepts, so we're interested, in trying to find out how to developers. On github rate concepts. On github so for, instance the word commit, means. Something different to a developer, than it does to someone.
Agreeing. To be married let's say and. These were rated by 50 github developers, and then we've done some work on trying to do, some sentiment analysis, so based on pull requests comments on github can, we extract these three-dimensional vectors, from that and. Then we have a few other data sets as well that I've been that I've been working with. So. The basic idea here of this project, is we're going to have all this data coming in we're going to use some NLP, to, try, to uncover what's, going on but, we're gonna guide that using, this app that control theory and hope, the two things meet in the middle and then, that will allow us to analyze, these different biases, that are present in these, social networks and possibly. Also allow us to design interventions. And then, artificial, agents that act in the networks in order to help the groups become, more more, effective, or. Have, fewer, biases, for example. Okay. So just, to sum up a lot of artificial intelligence is sort of I believe, sort of equated intelligence, with with, decision, theoretic rationality. But. We really know that emotions are necessary, for intelligence, and this. Low road gives heuristic. Social, intelligence and, that social intelligence allows, us to solve social, dilemmas, like, the tragedy of the Commons for, example, AI. May, be sort of pure desert traditionally. I people. Might say well you, know this, is just the reason that we have these emotions, as humans is just because we're, so dumb right there we just need to have bigger brains and that we could do everything decision. Theoretically, and we'd be we'd, be all set so. I think that that's fundamentally, incorrect, and. One. Of the ways to see this is by, looking at the different forms of commitment, that people make to one another and. This book by Lawler describes, these commitments as being instrumental so. These are things that are just have, to do with the, sort of practical, things like you know if I find you to create a coffee and I'm in a new town that, I i I just want to pay money to, a coffee shop and get my coffee I don't care which coffee shop it is if, one coffee shop shuts down I don't care I'll just go to the next one but, as soon as I walk into one coffee shop and I have a little interaction with the barista then. My, instrumental, commitment, turns into an effective, commitment, okay. So now I like, that's, my barista I like that guy he likes me next, time I go to the same two coffee shops I'd be like oh yeah I want to go to that one and then if that one shuts down I'm like oh no oh my heart is broken my favorite, barista is gone okay. These, affective, commitments, are much stronger, much, longer, lasting and run, much deeper they're, really what motivates, humans to interact. With one another so.
They Strengthen, the bonds that, make them longer-lasting, they. Also increase, trust, between a, and, this, leads to agents, being able to. You. Know interact with one another using shorter, contracts, right, I trust, this guy so I'm gonna do what he says that, allows, agent to be more efficient, and I, think this is regardless, of the available, cognitive, resources, right. So, a group that interacts, using, this is always, going to do better than, one that doesn't. Okay. So that's the end of it so. I. Know lunch is probably almost here, if I already here but are there any quick questions like four minutes for Jesse. Christmas. So. So, jesse has some free time on a schedule this afternoon between 1:00 and 3:00 so, if you do want to, you. Know book book some time to chat offline you. Talk. To me talk to him we can figure that home great. Thanks so much. All. Right.