Math Models Mind

00:00:00 yeah sad laughter still deep you monitor

00:00:34 my choice tasks and it's a question about some complexity creating models how people created models of the task is this can people hear me yep hi good

00:01:04 afternoon everybody I'm Gerald Horowitz I'm associate director here at the helix Center welcome to a very interesting roundtable today entitled math models mind I wanted to make quick announcement that coming up in future roundtables we have an March ninth life in the universe which we've established actually that right now today there is life in the universe but there's going to be more to say about that topic and we have

00:01:35 assembled the following roundtable participants Caleb Scharf astrophysicist at Columbia University Yale astrophysicists p.m. bada not Orion Kenneth dill chemist is Stony Brook University ed return a physicist from Princeton and Dennis over by a New York Times science correspondent so that's March 9th and I'm sure it's gonna be really quite a lively interesting talk

00:02:07 in April that I don't know if we have an exact date yet but I don't have it here we could have a discussion on climate change and the Anthropocene era which is among us which we're living through now and I'm May 18th you don't want to miss it or else you'll be in trouble we're having a talk on shame today today's panel I'll quickly introduce and you can raise your hand

00:02:38 when I mention your name first as Larry Ansel and Larry is a clinical research psychiatrist on the Faculty of Columbia University he's had a long background in mathematics and was an early proponent of using decision theory game theory and behavioral economics in psychiatric research I'm gonna keep this short so we have more time to talk Cheryl Corcoran is associate professor of psychiatry and program leader in psychosis risk at the

00:03:09 Icahn School of Medicine at Mount Sinai together with dr. chechi of IBM she's identified patterns of language that precede onset of psychosis including reduction in coherence and complexity of the speech andrew gerber is a psychologist psychiatrist and psychoanalyst who is now the president director at Silver Springs Hospital silver Hill Hospital you know of course

00:03:39 I was wondering where silver springs house Maryland that's what it wrote and that's what I read I'm sorry Ken Miller his professor department of neuroscience and Department of physiology and his director for the Center of theoretical neurobiology at Columbia University he's co-director of Columbia's source program and fear that remember algae its centre for theoretical neuroscience as well as its neurobiology and behavior graduate program and John Mary is assistant

00:04:15 professor of psychiatry neuroscience and physic that University's School of Medicine where he directs a research program in computational neuroscience with a focus on computational models of neuropsychiatric disorders he received his PhD in physics at Yale University or we have one missing number I'm going to give you his background in KC when he does show up it's our DJ rod nod he said associate professor in the mathematics department

00:04:47 at NYU and he's completed postdoctoral work at NYU his research has focused for many years on computational and theoretical models of sensory processing particularly vision and olfaction so thank you all and I thought because the topics here are somewhat stopped to give

00:05:22 authorities we just announced your thanks for joining us I was about to say that we were just getting started so the topic some of the topics related discuss today are little R King and I thought it would be helpful maybe give a brief summary of the kind of research they've done and/or they're interested in the topic of mathematics as a sort of modeling behavior and the

00:05:54 same so we're one side so what I've worked on is the the circuitry of cerebral cortex of sensory cerebral cortex primarily primary visual cortex but trying to understand how the circuits trying to understand general principles of how the circuits of cortex work and basically trying to understand what operation two circuits are doing to produce the responses to sensory stimuli that we see how those circuits develop through learning rules based on the

00:06:26 activity in the neurons and the faraway goal is to understand from that really what what computation cortex does the cortex is a very uniform structure half full half empty there's a lot of differences between the different areas but the first thing you notice is it's basically the same architecture with variations on the theme and so there's a sense that there's been a unit of mammalian intelligence that's been developed and then duplicated and applied to almost everything that we do and I'm I would someday I would like to

00:07:00 understand what exactly it does to the input it receives and Albert represents it how it transforms it how it learns from it so my research is also kind of focused on the article models of neural especially cortical systems as as Ken's we primarily focus on computations associated with a Association coordinates in contrast to sensory cortex looking at a kind of fundamental cut cognitive computations such as working memory or decision-making and then we're very interested in understanding how synaptic level

00:07:32 disruptions of those circuits could give rise to cognitive impairments as we see in psychiatric disorders such as schizophrenia and so we collaborate the experimental and study you know disease processes in pharmacology in these computational models of circuits and I I too have spent a few years looking at computational models of neural circuits in this case ice I've worked a little bit on envision but I spend most of my

00:08:03 last several years working on olfaction trying to understand the kinds of computations that go on inside the olfactory system of insects and as Ken was saying there there are many similarities between the olfactory systems of a lot of different animals and at first I was seduced by this thinking that perhaps there would be a similarity of a cup to the computations that these that these different olfactory systems perform however the more I look at it the more I've come to realize in recent years that it's almost

00:08:36 like there's a different different different operating systems running under the same way on the same hardware so you can talk more about that later don't left ok so so that sort of shifts a little bit I'll work back where it's my Curt role in some ways this as a translator between different languages around mental illness so in leading a psychiatric hospital my predominant interest in is how to get people to talk to one another not just between our staff and our patients or the patience

00:09:06 patience but between the staff and each other and one of the languages that that's been very important to me is the language of math and or more broadly speaking the language of modeling and I find someone to my dismay that that language isn't taught in most clinical programs and there's a way in which they're sort of urgency for an answer and I think this pervades maybe all of medicine but particularly in psychiatry and psychology makes people not

00:09:36 understand the value of models that are good but not perfect and one of the most common quotes I use which others may be familiar with is from George box who said all models are wrong but some are useful and to incorporate that is basically one of my main missions prior to being a president of a hospital I had run the MRI research program up at Columbia and so I was particularly interested in the use of structural functional MRI to test various mathematical models of neuro

00:10:07 cognitive processing exactly the stuff that you all are talking about I'm a believer that there are a set of neural cognitive processes that underlie our psychiatric disorders that actually will end up looking quite different from the symptomatic descriptions that we now use so the terms of depression anxiety even psychosis are appealed because they're very experienced near to the clinician but that the underlying there are cognitive vulnerabilities probably in my opinion that will look quite different and we're not quite at the point yet where we understand those so that's

00:10:38 that's been the sort of overriding theme of my research so I'm a psychiatrist and I collaborate with people who do computational modeling and I collaborate with key Gemma Chucky and his team at IBM and he has applied computational analyses to behaviors such as language so we think of language as really big data at the level of the individual languages semantics and syntax and there's pragmatics and probably facial

00:11:09 expression also has semantics and syntax as well as gesture and we believe this behavior in and of itself can be modeled and I think what we do as psychiatrists is we observe people I do think that data is very important and not only can computational scientists help us but they really feel that we can help them in terms of building the model so I've been to a few conferences the nips

00:11:39 conference nor informatics you should yoshua Bozzio is there about deep learning and someone in the audience asked him you know what what are your thoughts about what you could learn from you know cognitive neuroscience human neuroscience and he said I don't know and I don't care which is very interesting but our the the people that I'm working with who really do think about artificial intelligence a great deal feel that we have a lot that

00:12:10 we can teach them they really want to model what we do and the other thing I want to just say briefly is I looked at the kind of description of the history and it's you know has this feeling of kind of a guy on a hill right and the environment but if we think about behavior behavior at the level of milliseconds it's very interactive and so we want to model discourse as well in all of what anybody does here who's an

00:12:40 analyst or a therapist that kind of discourses is is kind of key to our therapeutics and we'd like to sort of understand what is going on in a successful interaction therapeutic interaction if that can be modeled by the computational scientists and then with that model use that to help you all in teaching other people what you do hi so um I'm Larry I am Sola and I got interested in this stuff it's 20 years

00:13:13 ago and actually got interested in it because it seems to me that within the social sciences the economics was the social science that most used mathematical mathematical model so I got very interested in in economic models so I think we're at work in this panel we're at wonderfully different levels some people are doing stuff neuron level with small circuits small circuit level and retirement all in instead of the patient and then turn words and the whole tape in the whole patient level but I think this stuff can be very abstract and so Jerry and I agreed that we allow me to give it

00:13:44 like a three minute talk on a trivial example and if I can just have you guys help me it hasn't been a save slides this is this line of the cygnets it's a it's a single slide but I actually need one back myself thank you so if we could get those moving around very quickly I would appreciate that so I would say that that what actually happened is that I was interested in I was watching as as the managed care movement was taking over medicine and they decided I needed to understand some economics so I went to

00:14:16 self of course in the economics of medicine and I bumped into decision theory and game theory and I realized that the economists really have some very interesting insights and very interesting insight into human behavior and the thing that that I had I had come from as I I dropped out of graduate school of math graduate school to go to go to medicine because I looked around the room and I realized I'm good at math but I'm not gonna have a job so you had to be great right so so so I got interested in these economic models and

00:14:48 then mack found myself working in a suicide research research lab and i started talking to people about the the architecture thinking like an economist what is the architecture of a suicide decision so if you look at this this is a very very simple basic model and the point of this is simply took to show a trivial example about how mathematic applications may change the way we think about something so in thinking about this the architecture of suicide sitting there's three smallest small assumptions

00:15:19 that I think people might agree with and that is that agents and that's another word for decision makers have preferences over orderings you know where you like vanilla better than chocolate you have preferences over outcomes there are outcomes in the world and you have preferences over them you like one thing rather than the other the second is that agents can choose actions but they can't choose the outcome so you can go and order the chocolate ice cream but you can't define whether it's gonna be good there's always a probability it may be good maybe not the target so I can choose actions but I can't always choose my

00:15:50 outcomes and the last one is that when I make my choices I try to maximize my preferences I do what I want to do rather than what I don't want to do it's very trivial and yet these three is very very I think easy assumptions that are not that hard to believe give rise to this decision tree which is that when someone's facing this decision they can either make a suicide attempt or not if they make no attempt and they remain in the status quo and that's the branch on top if they do make an attempt those are probabilistic and they could end up dead or they could end up surviving the attempt so there's three

00:16:22 possible outcomes in the decision architecture and here's where the trivial mathematics thing because if you tend to discrete mathematics on day one they will teach you that three things can be ordered in the six ways three things can be ordered in six ways chocolate ice cream vanilla ice cream strawberry ice cream there are six kinds of attitudes that you can have depending on on your watering so the mathematics forces me to believe that there are six type of clinical people six into six different types of people in their approach to suicide let's see if that's true well person number one is the healthy normal person is not suicidal he

00:16:54 prefers the status quo his second choice would be to survive a suicide attempt that he really doesn't want to die that is less that's his last choice on the on the opposite end is somebody who death is their first choice and they are willing to survive an attempt because the status quo is their absolute last choice and this is somebody who would make any let any level of lethality attempt they don't care about the probability they would make any live validate them the third type again this is purely mathematically driven the third type is someone who's surviving the attempt is their first choice the status quote is the second choice and

00:17:25 death is our last dress and this is what we used to call manipulative so is suicide but it to people who want for one reason or another to have made an attempt but to survive the attempt because that would either change them or change the environment that they're in we've all treated patients like this who suicide attempt I think these people will not make a high level suicide attempt they will only make a low level suicide attempt they're not a danger using a gun they're only a danger of low suicide attempt the fourth one here's somebody who is fear who really wants to take their lives let's say they're ill or

00:17:55 terminally ill people but there terrified right we're facing this so much they're terrified of making attempts it in the ending up worse right so this is the reason for the Hamlet Society and the reason for people like a vork Ian and it's the whole idea of euthanasia it's people who feel it's their time to die but they're terrified of making an unsuccessful attempt that's the fourth type the fifth type is somebody who their whole goal is to get away from the status quo of status quo is their last option and anything else is but it's better for them and finally the sixth type is a little bit hard it's

00:18:25 somebody for whom the status quo is fine but if they were to make an attempt they would not want to survive the attempt and that's a samurai or in Antwerp also has to do with dueling cultures it has to do with honour cultures now the point is that I I made three very simple assumptions that 90% of people I talk to agree which is it's trivial model and the trivial the trivial model because of one piece of mathematics that three things can be ordered six ways predict six types of suicide attitudes and most collisions I talked to say I recognize

00:18:56 those guys I recognize those people those are those are real categories of people and the talk that's applications mentioned just a few minutes ago from their sort of neural systems actually not just all social systems and people involved in a web of

00:19:28 relationships with others so I'm straining and think of a way to get ball rolling with a conversation because this is dealing there was a CRI guesses on mathematical models I thought it would way to get started we just throw out the following question I think a lot of people who are interested in mathematics and get into this field often find themselves and have been frustrated where pieces start leav lane they found themselves frustrated by the sort of lack of mathematical intuition the part

00:19:59 of a lot of people then you're the people who typically have done the kind of research you've done may not be so well-versed we're interested in mathematics so I guess I want to know how do think mathematics is an improvement over a similar approach to what you've been engaged in that does not involve math I mean I can jump in on that because I so I'm trying to write a popular book on this and it's very hard it's really very

00:20:29 difficult but I'm working on the book and it seems to me I was trying to understand what is the difference between a mathematical model and a verbal model I mean I grew up in the analytic world right and we had a lot of verbal models and I think the difference is that when you write a mathematical model you are committed to the consequences of that so that's that's kind of a thing there's an equation you are committed to the consequences of that of that equation where's verbal models and we saw this in poppers of critique of psychoanalysis right it was non falsifiable because you know the psychoanalytic statements were

00:21:01 sort of vague like what the consequences were I think but when you write down a mathematical model you are you're you are committing yourself to all the deductive and computational consequences of that model what I think that's different than verbal models but you guys would know better yeah what I would say is that by certainly for me working on circuits by exploring a mathematical model of a circuit you discover things that you would never get just by thinking about it and I think of a model as a scaffolding

00:21:34 that you use to develop new intuitions once you've got them you can apply them without the math but you'll never get them without the math and I can give an example in the in the brain and the cortex that but across the brain there tell that either that are excitatory they excite other neurons or that inhibitory they suppress the activity of other neurons so there's a there's a phenomenon called surround suppression where if you have a visual stimulus right where a particularly cell is looking it'll respond to that stimulus

00:22:04 but if you put other stuff outside of that region it'll tend to suppress the response to the center stimulus and so everybody imagined that the only the only long-range connections are excitatory so everybody imagined you're sending excitation to the local circuit so you must be exciting the inhibitory neurons and so they don't gets around suppress that gets around enhanced and then they suppress everybody else but then David first urged in an experiment that showed that when you when you add the surround the inhibition the cells received goes down

00:22:35 the excitation they receive also goes down and that's what's causing them to get suppressed and so then we had to figure out how can that happen how can you add excitation into a local circuit and make both the excitatory cells and inhibitory cells all get suppressed by adding excitation it turns out there's a really simple mathematical answer to that which we had to you know struggle for a little while to understand but then once you understand the mechanism then it you have a new understanding of how these things can work that you then it then makes predictions that you

00:23:05 didn't anticipate then experimental school art and test I mean I would say that you know it is the distinction between verbal models and mathematical models is more important than distinction between using models or not right and there's often the kind of a take of you're using a simplified model but the reality is that in science we're always using simplified models right if you don't have a mathematical model you have some verbal model some picture model you've got some arrows and this does that and this is how we kind of synthesize different facts plan new experiments etcetera etcetera right and

00:23:35 so there's always you know some kind of model that people are using or assuming even if they're not making it explicit right and so by using mathematical model you number one yeah commit to to make an argument for the sufficiency of something such and such mechanisms are sufficient to produce this phenomenon which is kind of all you can do with models really make kind of sufficiency argument so I would say as chemist saying we discover you know counter to it of things because there are you know you use the word kind of emergence that

00:24:06 when you have interacting elements in a complex way you can have resulting phenomena that emerge at a higher level that are not really you know clear it's hard to predict that from even the properties of the low-level property elements and you know I would say that a key role mathematical modeling is about bridging these different levels right where we understand something about you know take Ken's example about the synaptic properties and interactions of excitatory inhibitory you know that's at the real you know kind of cellular synaptic level and then

00:24:38 the phenomenon of surround suppression really emerges in the circuit at the physiological level and you know the it's through mathematical models that were able to show how the elements that the low-level produce the the phenomena at the higher level actually one other thought that occurs to me you know I you said the you know all models are wrong but some are useful would and when you said that it certainly what I actually think is all models are incomplete I don't think they're necessarily wrong like for example I don't think our model of how it is that

00:25:10 when you add excitation to the local circuit everybody it gets suppressed I don't think it's wrong but it's incredibly incomplete it's a very very very simplified model I think box was being intentionally provocative when he said it that way yeah yeah I think he would agree with you I would so my I thought about about your question though is is to try to disentangle a little bit than the notion of use of mathematical models from the personality of the person who's asking or applying the

00:25:41 model because I feel like these things get conflated often and and it's one of the struggles I think that work that exists in the clinical particularly in psychiatry and psychology world which is I think there's a disproportionate number of people who get attracted to clinical fields because of a pleasure in thinking in sort of almost rebellious ways but at anytime you hear something you sort of think well how's that also not true and how is that incomplete and what's being left out of that which is

00:26:11 which is a personality style it's personal I saw that I that I like that I think is very valuable in clinical work but it's interesting because it's often seen as a precise though that's at least in people who are non mathematical as being incompatible with mathematics and I don't think that's true at all I in fact think that that that when you have a set of rules when you have a set of of practices constrains you can actually think out of the box and even still more a provocative way and and I think

00:26:41 getting that's a nuanced what way of describing the use of math that I think many people don't get they assume that if you do math you're content with vast oversimplifications and it never occurs to you that there and I you know I see laughing because I've never met someone who's mathematically inclined who was content with vast difference no I should pick up on one of the things that Andrew said about character in the people who were attractive I mean one of the things that that it seems to me as I wondered as to why you know in psychiatry had not

00:27:13 become mathematical I mean Cardiology's Mathematica there's a lot of more mathematics and cardiology than there is in psychiatry and psychiatry about the brain so you would think and I think that you know was there was a famous book by CP snow called the two cultures talking about you know the time when there were there was a time when people knew everything and then there was a time where people sort of divided into the humanik quote unquote humanities in the sciences and I always wonder whether people who were drawn to psychiatry were sort of the same sort of people who were sort of thought you monistic aliy and that means that means that they like to

00:27:44 think in metaphors rather than in formal math models right that you think in terms of metaphors that you 13:13 similes and that i think that is the difference between sort of the humanities and the in the heart and the hard sciences and that psychiatrists were generally drawn from those kinds of people that i could be wrong what is a mentor if not a model you know and and so that's that's you know people make this argument frequently and and i agree with you as a cultural level and then i think why can't we get beyond that why can't we see that you can actually apply math to these kinds of

00:28:15 more humanities models without doing any damage there's what there's a wonderful quote in the beginning of Freud's bark field leonardo da vinci which which i always remember where he he says that some people have criticized him for embarking on a psychoanalytic study of Leonardo's life and those of you read it more recently than I could can correct me if I'm wrong here but he says he thinks exactly the opposite he thinks by applying a psychoanalytic understanding to Leonardo's life rather than oversimplifying or somehow reducing

00:28:46 Leonardo's brilliance he's actually expanding and making it even more exciting that Leonardo ended up as he did and that's the same relationship I think of math to them to models that come out humanities which is it actually doesn't reduce them but rather gives you the opportunity to eat them now and that synthesis is occurring in fact about two years ago someone came out with a again theoretical analysis of some classic pieces of literature of Jane Austen for example so it was a game theory model of Jane what one of Jane Austen's novels which is a really beautiful piece of the

00:29:17 kind of synthesis that you're talking about I think where I find the math you have to know be able to know if you're right or wrong if you come up with a pretty math model but you have no way of telling if you're right or wrong it doesn't get you very far although you know into something physicists debate now a super string theory but that's a whole other topic but at least for the rest of us it's it's important that you be able to you you come up with testable predictions that you wouldn't have

00:29:48 thought of otherwise and that they can be tested and without that you're kind of you have an insight that might be right might not be right and you don't know what to do with it and then I think the other element of being successful is really what John said is that no interactions of what what when I when I first was looking to do theory in biology I was if I was trained as a physicist but I was wanted to work in biology all the violence I talked with told me you can't the biology is a experimental science we just find out the facts and if you want to do theory

00:30:18 you should stay in physics and but why why neuroscience started to need theory was when we started to collect enough data at different levels that we had to know how the interactions at one level could create the the behavior at another level and that you those aren't just facts you can't just think your way through that and that's where you really and and then the field of theoretical neuroscience mathematical neuroscience has exploded since about the time I added the field happened to enter at a

00:30:49 very good time when several people were entering and that's when the field exploded but that's is it's yeah if you're not dealing with sorts you know how the rules of how a neurons activity evolves based on the input it receives excitatory inhibitory from other neurons how does that lead to this behavior which is more than the behavior you can get from any one neuron it's those kind of questions that where you really can't do it without math now I think there's a lot of this question of what of what does it mean in the mental health fields to be right or wrong

00:31:19 because because I think that that is a place that there's some vulnerability one can say what does it mean to be right or wrong about the ways a person that goes this person of depression functions and and I think what that part is what part of why I liked which box wrote so much because I don't know that we are at the stage now where we can we can we have ways of measuring right or wrong we have ways of measuring useful or not useful though and I think that one of the things mathematical models do in our field is generate new ideas generate new hypotheses that because one

00:31:52 of you were super saying one wouldn't have already one would one wouldn't have already just immediately concluded that emerged from the model and then you try it and sometimes it leads to something useful and sometimes it doesn't and one has to be willing to go down that path even with an absence of certainty yeah but that's what makes the models right or wrong is that they do make you know different to social predictions you can't test them yes or no right so I think we can distinguish that that that good modeling in neuroscience or behavior is kind of you know goal goal

00:32:24 oriented we're trying to explain a certain phenomenon we're trying to make some actual testable predictions and so that's in contrast to I think you know two approaches right one would be trying to say well I want to build a correct model I want as much detail as I can in there and have a complete model right and I think the point is about say you know a somewhat futile effort and not actually a good path to generating understanding or advancing you know how what kind of experiments what we should measure etc that's on one end and the other is on being kind of too abstract

00:32:55 too much of a toy problem which you know some people from from physics and mathematics have a tendency toward but but then there ends up being a gap between you have a toy model you can study some interesting properties but there's a real gap between how you can connect to that to experiment and to real data and so I think they're kind of two ends that you think fall off the edge in terms of being useful for modeling really just your clinical example so clinical example what you're saying in terms of how some of these mom actually being clinically applied so because of where we're located we're at

00:33:26 we had a huge 9/11 population of firemen and cops after not after 9/11 and I was in that I was in the trauma section and though in those years and you know we had a lot of I mean there are there are validated treatments for PTSD and we always use those kinds of things but what I found was that you know one of the things that drove the fireman was treating completely crazy because there's they're straightforward rational they're straightforward sort of very straightforward people the firemen are very straightforward people they're not very introspective who really bothered them that they knew that getting into an elevator was safe and they also knew

00:33:56 they couldn't get into an elevator and it was ruining their lives and lost their jobs over this and I found you simply take a reward learning model right and I said let me explain to you how this how this works your brain works in a reward learning model and you've had this one shot very negative learning thing in which getting is an elevator is that and it's it's not it's working in your amygdala it's not working in your frontal cortex so talking about it's not going to change that only behavior is going to change it and and giving them that model giving them that reward learning model really change things so it it made them feel less crazy

00:34:28 it made them feel hopeful and it made a huge difference does what's your your idea I just want to make one quick so maybe clarify and comment I hope is clarifying anyway Larry and Ken both refer to the idea that mathematics once you engage in the mathematical model you sort of committed yourself to the consequences of the mathematics of it and I think there could be this misunderstanding that means that mathematics is rigidly inflexible and so in and and therefore risks being reductionistic and missing something and

00:35:00 Andrew said something very interesting I think about real mathematicians in the real world or anyway in applied mathematics where the whole idea is okay if there's a remainder if there's a false I'm calling it a remainder if there's a false outcome or hypothesis that's been disproved put it back in a hopper let's let's use it and build another model based on that and that is a way to keep the mathematics from falling into being excessively rigid okay but having said that I want to ask the following question so there's a lot

00:35:31 of talk about whether or not can whether the brain is in computer okay now I want to have something slightly broader of course it's going to overlap with that question is the brain mathematical of course of course discussion about mathematical models with what is the difference between a mathematical model and careful explanation careful explanations don't go into equations which could be manipulated to give you other consequences we're not

00:36:04 well that doesn't become mathematical so my argument that you would be to say at the point where it's exact enough and exact enough I would say has two properties one is all the variables in you that you're referring to are measurable things in the universe agreed measurable things in the universe and the relationship that they have with each other in the universe corresponds to the relationship that they have in your mouth whether it's a verbal model or an equation model right so that there's a correspondence but if you've got a verbal model that has those two

00:36:35 qualities it's a mathematical model I mean laws are written in write or written in words but I I I'm with you in the sense that I think the word math gets used in a funny purpose here it uses it's used as a sort of synonym for for thinking carefully and and that gets very confusing I think you have to define what one means when one's having this baby I think the times you think you're thinking carefully and then when

00:37:06 someone forces you I mean this is sort of when you're doing applied math course right so you say oh I think I have I think I have a good model I think I'm thinking and then when you try to put it that in a mathematical way mathematics show I think the formalism I'm not disagreeing with you but I'm saying the formalism of mathematics does force you to make sure there's probably do that relationships carefully I'm an empiricist right I do science I don't do modelling to be what seems like the difference is that when you have a computational and mathematical

00:37:38 model you can look at latent variables that are not evident to the person doing impaired work and that that seems to me to be the difference you could argue second analysis was doing that from the beginning without using math but wasn't was often not doing it in a very precise way but oh I wasn't even thinking of psychoanalysis that sort of adds to it but I think within within sort of research you know I use statistics I can

00:38:08 use very advanced statistics but to model you have to it seems to me you're looking for latent relationships and variables that that's one of one thing is that you you look for you know a much lower dimensional set of variables that explain some high dimensional data but another is that you're just looking more like a dynamical system that's just understanding how the dynamics of the interactions lead to certain results in this but I wouldn't call that latent

00:38:38 variables that's sort of a just a different thing that makes sense let me ask the question different way cuz I've got such a Smackdown for the last the brain cannot be accessed through method that's a different way to ask that I think the same question well yeah so here's an example so I I don't know if everyone is familiar with the recent

00:39:10 advances in artificial intelligence through what's called deep nets but these are very very very loosely modeled on neurons with the coal neural networks and by being deep you just mean you have layers and layers of them so this this one predicted - this one was predicted this one was predicted - this one and that program has been going on for a long time and it almost died there were a few few Hardy souls who kept it alive through period winner most of the field didn't believe in it and then they had a

00:39:42 spectacular breakthrough about six years ago seven years ago and now it's completely taken over every field of artificial intelligence so Geoff Hinton who had was really yes sir one of a few leaders but to my mind the leader in the field - all years brilliant guy and and he's the one actually his group made the breakthrough that broke everything open in 2012 but he was asked so there's a big question with the deep notes which is that they have like a billion parameters a billion

00:40:14 synaptic weights that that have to be learned from by just knowing for every input what output you want and then you have an algorithm to make to learn those billion parameters to make this input produce that output and he was so there's a big problem is is there a black box we want to we want to know what they're doing we want to know how they decide that this is a car and this is a German Shepherd and we don't it's just a black box but it works really well and so somebody was asking Jeff you

00:40:46 know are we gonna be able to understand these things that he said no he said these involve a you know a billion parameters if it was reducible to some nice simple operation the problem would have been solved a long time ago people tried to solve it by using nice simple operations they didn't get anywhere and so that's an example of a level at which no you can't reduce it to any nice method I mean it is mathematics you know that it's a mathematical set of equations that are these these deepness but we can't it's not understandable

00:41:17 mathematics it's understandable at the level of what you're trying to optimize and what your learning rules are but not what it's doing when you presented input at the bottom and it works its way to the top that part doesn't seem to be understandable the key thing that's using that example its consent some of the other parts of that are understandable right the learning rule is you know one equation the loss function this one equation what is it trying to optimize and so on what in one sense it's a very simple description of what goes into the model but again you're training the sleep network to classify you know visual

00:41:49 images and so all of the complexity of the visual world of all those images is being you know parsed and extracted and that goes into those billion parameters and so in some sense the model has to be complex because it's ant processing complex data and that goes into the parameters and so in some sense you know we understand something out those models what are they but the final solution is very complex and I think a lot of neuroscience will will be the same way right but it's not even so different from physics right we

00:42:19 understand you know Newton's laws but you know we can't predict the weather right we can't predict a hurricane there's yeah but we still think the same basic fundamental interactions are there but it's a level of complexity that we're never going to understand on the same intuitive level as you know in a simple inelastic collision of two particles and so in the same way in in neuroscience in psychology we often will build kind of you know simplified models or we want to understand and understand some principles and then as we're saying it goes back to metaphor but often our metaphors are from you know simple

00:42:51 models are very constrained behaviors that we can model that we can study that we can you know look at in humans in in animals etc and then there's this process of extrapolation and metaphor to the kind of full richness of human behavior yeah I think a clarification and it was we real actually worried about this coming into this talk as the clarification is that and I always worry about the term computational psychiatry I used to call it mathematical when I started doing it because what I was doing was math modeling right now a model a math model of behavior but computational concrete encompasses two

00:43:22 things that are actually at the polar opposites one in which you have specific models like a specific reward learning model for the way a mouse is gonna is gonna behave in a maze and there you have very specific models but then you have these computational things which are absolutely black boxes so it encompasses both these things in which not only are there no mathematical models for the beginning to end thing it's not even possible that will never be possible as as you point out because you know deep networks will never be reducible that's why they work it seems

00:43:53 to me that the that the subtlety you're talking about though between essentially the levels of complexity and how much you can simplify these models is different than what people mean about is it mathematical or not from that question at least the way I understood your question it's all math that's all mathematical yes how many parameters you have yes to what extent is it computational versus solvable into a simple equation that's different but but I heard your question and tell me as being more connected to the question of are we do lists or not because true

00:44:23 dualism and Larry and I've had this debate just recently to my mind suggests that there is an aspect of the universe that is does not obey these these same laws that is not quote-unquote reducible to the sort of principles by which we use now I am NOT a duelist I've become so I may be misrepresenting but if you're if you if you don't believe in dualism if you really believe in monism that there's one universe and there's one set of principles that governs it

00:44:55 then it's hard for me then I don't think that there's anything that isn't at some point amenable to the use of math I think that I don't know anything about philosophy but to me it seems very obvious that I feel a certain way being a conscious thing I don't think that if you make the assumption that that consciousness is some emergent property

00:45:26 of a complex system it seems plausible I don't think mathematics as it currently stands has any any foothold on on that question if there's there's no language that mathematicians have access to today that can begin to untangle why it is that I feel the way that I do about anything I mean and you just to answer me to push this analogy a little bit further you assume that we're conscious to whatever degree and other

00:45:57 similar complicated systems that somehow perpetuate themselves are also conscious maybe bacteria or sharks or dogs or cities or the stock exchange if there's consciousness in these large complicated systems I have no idea how to address it and I don't think any mathematician I don't think math if you say careful thinking will eventually understand things are short but if you look at the current set of tools that mathematicians have access to no way but actually I

00:46:30 think I agree with both of you and because I I think that that it I mean what you're saying is basically look at the physical system that obeys the laws of physics and in that sense you know you can describe it in principle with those laws you can't in practice because it's too complicated but it you can describe bits and pieces of it but it is it is a physical system that operates according to some actual laws and isn't just just doing whatever it feels like doing at any moment but

00:47:01 there are properties that emerge from that system right that are very hard to describe one you know when the physicist describes the properties of a water and why it's wet you know in terms of the atoms in the water what you can you can derive from statistical physics the attributes that you can rather recognize us as how wet things behave but so there you can actually see the emergence consciousness I tend to think that why

00:47:32 material things have a subjective experience is not a scientifically addressable question - why but the what you know which material things end up creating subjective experience you know with the correlation between the material things and the subjective experience I think will be able to you know in principle describe that perfectly so why you know why isn't a zombie why doesn't have significance perience I think that's not a scientific question I don't know how science could possibly address that so Andrew and I

00:48:03 have had this ongoing and I would say I'm a duelist all the way down which is to say you know like the turtles all the way down we should say I don't think you have to I don't think I agree with your art with your argument but I don't think I need human consciousness to do it I think if you if you simply look at a single-celled protozoan that if you put them in a in a protozoa feeis if you put it in a in a in a petri dish it will it will swim tore up this sugar gradient right and so that protozoa swims up the sugar gradient now I can describe the

00:48:35 mathematics of the turning of its you know you know the it's turning in its tail and how it turns and I can give it perfectly probably get a perfectly good description of how that but that will never replace the description that the protozoa that that that creature is swimming right towards the sugar right and those so the mathematical description will never replace the same the intuition and the description of what's going on and and I think those are the two levels and I think that's exactly what you're saying so it's not even only the human being

00:49:06 consciousness level I don't think science can ever I don't think mathematical things can ever they can explain the mechanism of it but you lose something when you move away from the thing of oh what's going on oh it's swimming towards the sugar are you missing something as in a description of observing this if you leave out the idea that it's it's swimming up a gradient and that swimming up a gradient wouldn't be in the mathematics sure could be something at every model depends what the questions are but you you can model

00:49:36 those behaviors this is what mathematical psychology does right you're trying to explain the pattern of behavior in relationship to the stimuli and you have real kind of you know they're truly mechanistic they're generative models of but of psychological processes not of neurons right and those are perfectly good quantitative mathematical models meant to address you know kind of again the level of behavior in relationship to stimuli or learning right and that's again bridging those two levels of analysis and then there's the challenge not of how does a model connect to different levels how do we connect to different modeling levels of analyses

00:50:08 the models of the level of circuits versus models to the level this isn't yeah I think it's debated incredibly clinically relevant because back back to your original scenario the typical occurrence is a patient comes to see me or one of us and says I'm having a subjective experience a deeply subjective of conscious experience that I don't understand they say that that that that is upsetting to them in some ways whether they're anxious or they're angry or they're sad and there's there's a question on the table as to whether any

00:50:40 kind of and I'm gonna use the word analytic not in a psychoanalytic way but any kind of mathematical or analytic thinking has any value maybe there's nothing I can offer that that's one hypothesis that patient because what they are coming to me with is so intrinsically subjective and doesn't file away a standard set of rules that no amount of experience I have could ever really impact I don't believe that I believe that using my experience using the models that I have about why some people get sad why some people get angry why some people get anxious actually has a

00:51:13 value for them not just in a kind of oh let me tell you a nice story and therefore you'll feel better but rather in that you'll learn something about your subjective experience that is going to be useful to you and so that's to me what why why there's a bridge between those two you know are there are aspects of subjective experience we don't understand yet absolutely most understand but some of it we understand in small ways and I think that's proven to be very useful and I think that's gonna continue to grow so this is why in you know clinical psychiatric research

00:51:44 there's such a push to go beyond kind of subjective evaluation toward actual you know quantitative behavior so I can we relate subjective experiences of anhedonia you know to the parameters of sensitivity to reward and Punishment in a reinforcement learning model that we can actually set you know can we relate delusions in schizophrenia you know does that can we encapsulate that in a more you know quantitative mathematical formalism of Bayesian inference and how we bring prior information in combination with with sensory evidence

00:52:15 right and if we can do that then we have it's there's still that metaphor that jumped between the simple computation and the full subjective experience but now we can study that quantify it study that in animals understand what the neural correlates are and that gives us I think a foothold in terms of bridging these levels of analyses which ultimately we need to in psychiatry if we're going to talk about how pharmacology can affect brain circuits and ultimately alleviate symptoms

00:52:45 computers so if I were to go head-to-head with your model and evaluating people for delusions I would win sure so what am i doing that is not in your model there is that neural networks are starting to get better at humans at a lot of towns like human networks again neural network better than humans that a lot of things

00:53:16 that involve complicated human judgment like what radiology well yes go tell me and go okay know what what they need right now to work is they need a huge training database but if you had a training database of you know the the conclusions you want to reach and the information that goes into it and you could you had a lot of that a machine might learn to do it better than you so when I and all of you look at a people

00:53:48 look at somebody and their facial expression you don't instantaneous calculation that's better than any computer so far yeah no I know they're learning recognizing motion stupid they're not as good at it yet but but there were all sorts of nuances and and then in terms of context I mean computationally we are very complex we are but we to build machines that are equally compliment and also I would point out that that by virtue of

00:54:19 evolution you've had a train you have been exposed to training sets right over a billion years right and a billion years of training sets have selected you to be the best facial recognizer right and the people who are is not as good facial recognizes no no evolution has been a training set over billion years we sometimes discount that that you know no but evolution has acted on the genes right right that has selected for the genes that can be to this computational ability you know not Lamarckian even though you know eugene epigenetics right

00:54:53 it's a little bit Lamarckian that's right but I have had several years of pattern recognition but that even really thinking about it and I'm awfully good at it and I I data set of from yeah in an interactive way close loop interactions will we be replaced where we one essence be replaced by psychiatrists maybe by computers it's not that you know humans are gonna be replaced soon because the only things

00:55:25 we really know how to do in neural networks right now are when we have a huge database of input to output then that we can get the machines to learn about better than us Jimmy these neural networks have never learned the damn thing they're trained to perform a classification task they're not learning anything well define what's your definition of learning they have no value this goes to the Chinese room argument I just wanted to finish that the point is that some things that we

00:55:58 think of as our unique human complex intelligence are are starting to be done by machines better than us like plane go like plane jizz and like like some kinds of recognition of objects and so forth and I'm not saying that it'll all be replaced I'm just trying to say that there's not it's not obvious that there's something in us that is so special that we can't someday get machines to do it better I agree that's an example with this question and it also shows how a morally complicated this is so I think you would agree I

00:56:29 think you would agree too Larry that one of the things that we wish we were better at as clinicians is predicting who will commit suicide that we do our best and when we work in the emergency room or work on the inpatient unit we make those decisions all the time as to who to keep involuntarily and who not to and and not infrequently yes wrong so let me finish the story yeah I have a friend who works at KU he says that it is as kind of an openly known at Google that there are algorithms that they can apply that enable them to predict who's

00:57:02 gonna kill themselves tewi frightening Lee precise degree the problem they have enormous datasets they have enormous access to data that we conditions do not and then the question becomes does Google want to admit that they know that or not right Facebook's now started to admit that and doing little things here and there is it enough I'm not sure but what do we as a society do and the danger to me of too much you know prioritizing of the clinical superiority would be that we might not take advantage of some

00:57:32 these other data sources that could help us say I don't want to pit myself against computers I meant that to be provocative I've done research in this area right so I've done research that shows that again in collaboration with Guillermo that IBM Watson when it looks at language patterns and individuals at risk for psychosis of whom 20% developed psychosis in two years does better than me I mean I've published that in terms of suicide there's a it's I'm sure

00:58:03 Google has much more than what's in the literature because they have a lot of resources but in the literature that there is a good literature that if you look at language and its across its people texting talking all kinds of things I reviewed this for a grant that I put in and Guillermo did an article actually looking at poems poets who committed suicide and those who didn't and looked at the actual language in their poems as to what would predict consistently across the board it's using

00:58:33 words that have semantic content but that has similarity with hopelessness and depression so these are tools I mean I don't want to put all of us out of a job but I just brought that up in a sort of provocative way in terms of the models because we're also we're also models right doing computations we don't understand their own computations but exactly I'm curious if you could expand on your idea about learning and whine neural nets don't learn um so even these these words I feel learning training etc

00:59:04 even even those are very loaded as you said before these people nets are complicated functions with many parameters hyper parameters even if you take into account the network architecture and there are objective functions that are set up which are minimized to fit the parameters of that function so that at the end the function has certain specific input output properties okay it's tempting to take a look at the goal what is called what was

00:59:36 the goal thing alphago alphago or deep blue and say wow it knows how to play chess it knows how to play go it's a giant lookup table that has been as their concern it has been pruned and refined over many many many iterations given lots of data if you sat alphago down on the table opposite whoever the gentleman was that alphago was playing you said okay guys new rule if you go ahead off the left side of the board you

01:00:06 come on the right side of the board you're playing on a cylinder now or torus alphago wouldn't even be able to take the first move its database would be inapplicable you sit down Garry Kasparov against deep blue and you say okay guys new rule the Queen by the way it's no longer Queen it's like the rook and a bishop I'm sorry we're gonna night it's like a working tonight but doesn't move like a bishop anymore okay garry kasparov would still kick my ass but deep blue wouldn't be able to do anything and so I'm not

01:00:37 saying that there won't come a time where the types of functions and the types of parameters that people are able to bake into these neural network just officiated enough that maybe there are some parameters that can be left to be you know tuned on the fly after the machine sits down and sees it but right now we're definitely definite there so so so um this issue is actually the philosophical issue this is philosophical debate about something called the Chinese room problem I don't

01:01:07 wanna get into a bit it's exactly this this point as to what it whether it's something why not argument is if I have someone you know if I have someone in a room who has who has well this right you know the you know the argument right and he has all of these dictionaries China saying Chinese English dictionaries in the room and if I give him a sentence written in English he can transfer right he can then look everything up and he can then respond and he respond properly in in Chinese yet there's nobody who

01:01:39 knows Chinese so that's the Chinese room argument there's nobody who knows China and he can the two sides of the argument are one side says that proves what your point is that the machine doesn't know anything and the other side of the argument says no the entire system knows Chinese and that's what we mean by knowing Chinese no no no it applies to the Chinese room because if you it's apocryphal right all those stories about like the the the spirit the flesh is weak but the spirit is strong right going into Russian and then being back translated as the meat is

01:02:13 rotten and the alcohol is bad right so it's it doesn't work that way with language so whatever he came up with without knowing it when you put a person in there they have a flexibility that the Chinese remark with the person is really acting just getting into a strictly rule-based language if you believe in Chomsky's sort of hierarchy now you're at the bottom and Chomsky

01:02:44 would say for like a human sophisticated grammar you need you can't it has to be probabilistic it can't be rule-based you could build all those things but it's complicated and that's the Turing test the Chinese room is the Turing test but for the champion Chinese room probably captures tries to capture this difference between what what what do we mean what is our intuition when we say somebody knows something or somebody has learned something and it challenges that thing and it challenges as you did this question of whether we do that's the purpose of the argument is to challenge

01:03:15 the the you know the use of that meaning just one other point that I run into whenever I try to talk about it and that is but you know we all have folk science in addition to a formal Sciences folk philosophy that you know cultures have folk philosophy we have folk chemistry and things like that and they're not very good I mean people have survived before formal science came in because they had these theories but folk psychology is really quite good right pre scientific psychology is actually much better we're pretty good psychologists just intuitively we have to be right we've evolved to be pretty

01:03:46 good psychologists and I think that is one of the reasons a lot of these things get confused is because because folk psychology is so good right it is hard for scientific psychology to sometimes compete and there's there's a much more sophisticated language within folk psychology than there is for example in folks physics or in folks right folk notions of mathematics and that that gets in the way I want to just respond briefly to Audie I agree with your criticism that there they they have a limited range and if

01:04:18 you if you change the rules on them they don't they haven't been they haven't up in talked to a debt but it's not to say that they couldn't be taught to adapt but the other the other thing is that I don't think it's right to say they're a lookup table because I also go well on any given game we'll see a board for that you can't see a board position didn't know him and has ever seen before and that the machine has never seen before and it'll still I'll play Lisa dough because it's learned somehow implicit in its billion parameters its

01:04:48 learned some principles of how to play go that it can it can deal with entirely new situations and I'll play any human and the same way that with alpha chess or whatever they call it your alpha zero yeah so I don't think it's right I think deep blue was basically a lookup table but I think so they're probabilistic now well no but the point is it really understands some principles in a way that I mean deep blue just did a deep search and said you know if I could look ahead ten moves if I do this am I gonna be better off or

01:05:19 worse off and it didn't know anything really what what do we evaluate better off or worse off was really simple but have this powerful ability to search but that's not what these new things are doing they are they don't search they learn nearly so deeply they learn principles that they can now play any human philosophical push back to you would be when you say it understands right well I understand I said their principles open these billion parameters read can now I'll play any humans but those principles as I understand it also kind of you know Gestalt sits observing

01:05:50 a pattern and then getting some kind of you know value function or if it's gonna play here not they're not necessarily you know rule-based articulable or ticular Bowl I can't even articulate it you know decisions but but basically kind of Pat it's a you know a form of pattern recognition but you know that also underlies a lot of our cognition the way we recognize emotions the way you know ain't probably a good chess player can just kind of get a flash of the chess board and get a sense of is this a good game to play or not this is a good move or not in that kind of intuitive way a lot of our cognition we

01:06:21 can't necessarily explain but so they're mathematical models for Gestalt a kind of Gestalt perception that we're all so good at as people I mean I would say a lot of the the visual recognition in the same way that people use you know deep networks for this you know we can argue about whether it's kind of training or learning about the same rules but you know as a metaphor of the general principle of you know how does our eventual ventral visual system work for object recognition of it being kind of successive you know layers of

01:06:52 feed-forward processing with some kind of flexible learning such that the representations at the intermediate layers of your vision you know are guided by the final you know output recognition and decision in action is you know I think where you would say that model is in some sense a good model of human vision do we already have our own neural net for yes only still very impoverished and now let you do what

01:07:22 they do but already just with that impoverish analogy you can outdo us on some test we could sometimes outdo ourselves in tasks that we don't think we're proficient at right those are examples of people being able to do fairly complex thinking without realizing in advance they know how to do it so you could consider that analogy to what some of these computers are doing perhaps yeah but you can also bring these back into cognitive neuroscience right so these deep deep networks not only do they you know do object recognition it seems like they do it in

01:07:54 a similar way as us so we can do you know recording from single neurons and monkeys we can do functional MRI in humans and we can look at what kind of neural representations that are in say different layers and visual processing and compare that to the different layers of processing in these deep networks and there are correspondences and so that's you know it's concet they're trained an impoverished way they're not little acting some things but where we can actually try to check the actual internal mechanisms they seem to have some validity in terms of explaining not

01:08:25 only our behavior but even how the brain works our representation I I think I would definitely agree that there is some analogy between what we're doing and what strategies are used to make these deep neural network that's true but but I also agree that it is impoverished and not but the jury's point about our doing computations that we don't know I mean if as anybody I've had the experience of these phony rocks that are really very liked but they think you know you look at the it looks like a rock but it's actually foam and when you try to pick that up you completely miss right you completely

01:08:56 miss the motoric thing because we take absolutely for granted that the incredible computation that goes into a simple motor thing of lifting something up and putting the right pressure on it and I can fool you by giving you an a Mis estimation of what the weight of the thing is but so we're doing incredibly complex computations right just for that simple motor motor action we have to plan it very carefully and we have to know what it do I give credit where credit's due I think that's one of the early contributions of psychoanalysis because I think for anyone else is not

01:09:27 on motor movement but to acknowledge that the complexity of these models outside of awareness as I think that's that some of the literature that was early literature on what was outside of awareness was and it had to be incredibly simple right and that more complex processes therefore had to be conscious and we now know and I don't even think it's controversial magog of neuroscience world that the kinds of processing that can go on completely outside of awareness can be enormously complex lutely and this was something that before just

01:09:58 talking about you know 100 more years and that example highlights one of those key computations is prediction right and so that actually gives us a hugely rich dataset because we're not just you know training the human brain on you know discrete supervised learning right but on predicting victim the future what are the properties of things in the service of predicting how I will interact with them how they will move in the future it's up

01:10:37 [Music] sir please have a seat operator we have an opportunity later to speak okay Larry mentioned earlier that you know in mathematical models we have these variables that are all well defined in which we can then check on or be aware of and the interesting thing is that when there are all these hidden units and in some forms of mathematics as well

01:11:08 you can only say that theoretically they can be checked on theoretically what theoretically they give me that they are possibly accessible to be yeah access but in in fact in the process of doing a computation some of those things are just there they're X's and Y's they're not really specified right I mean I guess I'm a distinction okay me neither okay no I think you reported by D in the deep neural net yeah maybe you may

01:11:39 understand you may understand individual weights but you don't actually understand how the process is leading to the proper route so that though that is it absolutely here there's an intuition of black box that's what I can refer to as a black box that there's an intuition we have about what it means to really learn something that I think does involve our knowing the steps quite well and then having an intuition that we know those steps that's what we call I'll say for the purpose of this comment that's learning but but but when they're invisible and they're being manipulated

01:12:10 all the while we then have this thought well maybe you don't know that's how we might be able to do things unconsciously well there's not aware of an area's being but I think that sense that we know what we're doing is an illusion because when when we've when I mean artificial intelligence for many years until the deep night revolutions about six or seven years ago was focused on trying to do things by writing down the rules trying to do vision by writing down the rules trying to play chess or go by writing down the rules and it failed miserably because we don't know

01:12:42 the rules the rules and actually to go to Jeff and it's point you can't reduce it to the rules it's a way more complicated than that I was trying to account for why we might have an intuition of what how learning is different from what computers do so whether or not it's good at it or it's accurate that may be what gives us the feeling whether intuition that we've learned something as opposed to just following a recipe I think I learned something when I can teach somebody else and I used I don't know if there's an analogy for these nuts baby

01:13:14 my dysentery they can feel they can teach other nuts it like I make a bad analogy which is that the people can take a neuron that this limb to do something and then it can teach another simpler net to do it it can teach it it can teach something else to do it more compactly teach it in the sense of being the being the one who tells it what the right answer is so I think people do use kind of one net to teach another in very simple ways right now I think that goes a nice don't really interact those just as they're trying to get out when I

01:13:46 talked about the folks I caught the folk psychological thing which is to say that we have this whole language that is that that we have evolved and it's about it's about ourselves understanding ourselves and sending others a folk psychology and then we have these words and then we apply these words and then those words have very specific human meanings so where I think you're right there when we say learn we got to be careful we have we should have two words right we shouldn't say a machine learns because learn very specific meaning it has it right it has it very specifically using I mean is you wanna stick meaning it has

01:14:17 to do with experience and it has to do with feeling and so I think I think that we applied the wording we apply the word incorrectly because we take folk psychology which was developed for human beings then we apply to machine to machines and those kinds of things and we don't have other words but we really would have other words we do qualify kind of you know there's reinforcement learning and unsupervised learning and supervised learning four different ways that we learn and again your reinforcement learning like you said may

01:14:47 engage the amygdala in unconscious way which may be different than right you know supervised training as well as just kind of general experience for prediction which maybe you're less conscious of sway but it's interesting to me how many conversations I think in a way devolve into arguments around the sum antics in exactly the way you're describing between the folk psychological use of a term at which people didn't want to defend in a certain way and then the other uses which may be narrower broader but are carefully defined and and and I think

01:15:18 ultimately since we're not going to discourage people from using those words you just have to get people to define them when they're using them because people use the same words in different ways I mean psychoanalysis hasn't had this problem for a long long time now each as we all talk about transference we all talk about libido and we can have very long conversations meeting entirely different things we've with the same word that then often can be unproductive Cheryl mentioned the Turing test which was the the definition of intelligence right so that was a debates you know forty years ago and debates on what does

01:15:48 it mean to be intelligent and they tried to define that the Turing test is I think is alive and well the term is the idea of the Turing test I think is alive and well aligned well maybe you should tell people what it is yeah oh so Alan Turing said that had a test that for artificial intelligence that if a computer could pass as human in

01:16:18 conversation with another with the human it will have passed the Turing test so you know you can think chatbots they can as a simulation for a limited amount of time seem almost human but if you really push we don't have any machines yet that really passed the test my way I don't think my father could pass the Terry if I talked to your father I wouldn't

01:16:50 think of his human with him via text or email I think you would think he was a chatbot be my the chat bot but it's not part of the Turing test the Turing test was before chat BOTS existed what I'm saying is the chat bot is getting close in some circumstances to appearing human I see no once no one has that where does the status of that it's not no one no

01:17:21 machine has passed the doing that no I don't think so I think there was one chat bot which kind of passed by acting as if it were a child who's a non-native English speaker been such a wonderful and lively conversation and invite people to ask questions why'd you come up to the microphone I just thank you

01:17:57 for this engaging enlivening discussion it's interesting because when you were talking about applying mathematical models to psychological states let's say it seemed as though you were mainly talking about behavior decision-making in action and so I could see how models of probability could determine you know could be used to determine particular outcomes and so the same thing when

01:18:29 you're talking about the the computers playing chess right that's how do they make decisions how do they act how do they behave what information do they have access to but what about mathematical models apply to let's say experience how one experiences or to consciousness in general that seems to be a little bit different than than the direction you're heading with like art young yeah how one experiences are to what creation is mathematics

01:18:59 yeah exactly but but even but even experienced it right how do I know what I experienced and what do I experience and and is that something that can be described mathematically given you have enough information and data let's say I think right now we couldn't say anything about that in principle when we have you know way more ability to watch what's going on and in brains and to you know

01:19:31 at some point we're gonna be other as I said before some point we're gonna be able to say these kinds of neural patterns of activity reach consciousness and and you know lead to this feeling or this or this experience and these other kinds of neural activity don't go to consciousness but you know why why they have conscious experience I don't know what we'll be able to make an isomorphism between you know the behavior of the neuron tend to be and what it feels with different things feel like but but that's way far in the

01:20:03 science fiction future where we're just we couldn't even touch that now we'd say that the why question depending on how you define it sure I can see that being far in the future what I don't think is far in the future is models that are relevant to effective experience and and you know Charles work is one example of that but there's many other examples now of various relatively simple models that look at the the sort of dynamics of various affective states and why one react certain ways to different things

01:20:34 and the correspondence of that between you know facial expressions and social papers but so I agree that's a different level I guess what I was addressing was its owner can we talk about feeling you know that's way far enough yeah yeah but that's also when you discussing modeling affective states that's also what gives rise to the affective state or the or the behavior of the affective state but not the experience of the affective

01:21:04 that's what I was drawing this depiction with the why I think one can always carve out a a realm of the Y that we don't have access to but I think that's shrinking I mean uh the example I'll give of that is I had a classmate Natalie training home remain nameless that who we used to have this debate about whether there was possible or worthwhile to measure anything in the psychoanalytic process and so he told me that his definition of psychoanalysis was that which I could never measure so by that definition as I

01:21:36 got better and measured meant better and better at measuring things his death his psychosis got smaller and small he was okay with that Houston but but I guess I would I would argue is is that sure one can always carve out an a Y that is inaccessible but I think it's relevant to the Y the more and more we know about about the dynamics of these ethics or the what absolutely yeah what but the philosophy philosophers talk about the hard problem my family with that problem and I think I think that remains obviously an open question and I don't

01:22:06 know I don't know that whether that's a scientific question or remains forever a philosophical question as to how you can bridge right between mechanism and experience I think that the hard problem is a hard problem for good for a good reason and I think it's an empirical it's a practice I guess it's an empirical point as to whether we'll ever be able to bridge that I'm betting though you can store can i yeah I just go back to an example you said earlier just to see where we stand on things let's take your protozoa in the dish going up the criminal gradient as Larry

01:22:37 pointed out you can have sorry Ken I'm sorry you can print it out you can I thought you can earlier though you can write a complicated differential equation that has every atom neuron whatever in that worm and your model will climb up the chemical gradient and you can point to every little bit in your model and explain in a particular way why that thing is climbing the chemical reading but would you ever understand how it feels how experience

01:23:10 is climbing up that chemical gradient and you say no that's what you say well I will say no when I was saying there are different languages because by the way when you say that something is climbing up a chemical gradient you're giving a teleological explanation right you're Blainey right you're explaining something by the end right and classic Aristotelian teleological model which is not a scientific model you can't describe something that you can't poss wait a cause that happens later right that it's climbing up the thing in order to get sugar that's a reason and can never be a mechanism I mean philosophically so that well but

01:23:42 but you can certainly you can certainly build a model of things that follow gradients and what how they behave I mean and you could model it at that low when I described it it's doing this for this purpose I've gone the conciliar mechanistic is those internal states about the past histories you can tell whether you're you know increasing or decreasing on the gradient and so even in this case for subjective affective States people have done things that you know within again the limited context of sequential you know learning and decision-making and reward you can you

01:24:12 know even say how does your effective rating of happiness correlate with rewards and reward prediction errors right with humor I think there is some work I'm seeing you know in simplified settings can we show how humor relates to you know violations of predictions of it's really pretty subjective but what are the kind of computations right at the level of you know inputs and outputs that's right you can relate these things

01:24:42 to actual computations that guide behavior in a mathematical way I would say so forgive me you could humor you can find some rules yeah but then to try to create humor using those rules that's where it hits a catastrophe or the model of humor which actually works pretty well I think that is something that humans we have reasons for doing things that that so far have not been built into sort of computational no no no you can want to

01:25:13 let us foolish agents that have goals and and and that have some relationship between their actions and their goals and their outcomes and they learn from that you can mop it in the context of evolution to what's selected it also said that we have good evidence that there are times when we think we are motivated to do things and that there's good evidence to suggest that we have told ourselves a story about wanting to do something that we were doing for completely other reasons and habits that are introduced oh no I

01:25:46 knew should I work with Cheryl I'm a neurologist so from from my perspective there has to be sort of an answer to all of this I guess my question is you know one of the limitations I see with with models and computers is that they have to be created by humans right so is it conceivable that perhaps the reason some of their limitations are because we as humans are not smart enough at least now to really create these models and perhaps with evolution or with human learning we could learn to create these

01:26:17 models and computers that could that could account for more complex psychiatric functions well from the point of view of the history of science the history of science is kind of on your side because there were all kinds of things that people said will never be explainable by science including yuria right where the production of urea and the in the human body was something that people thought couldn't be it couldn't be replicated so the history of science is sort of pushes to that but I would say the that that you know the human brain is the final frontier and and we

01:26:48 don't know whether whether our historical experience with science being able to progress will be that would be a limitation of us in our ability to create the model as opposed to a limitation to models in general couldn't you argue that one of the ways of thinking about our official intelligence is that it is teaching a machine to create new models when I could imagine certainly and maybe this is already happening building a machine that builds new models that we ourselves as humans couldn't build no it has to appreciate

01:27:20 art and tell jokes and it has to be human that's the bonus I don't I wouldn't my argument that there's some intrinsic limitation to how complicated a thing humans can create I you know it's just a matter of just the whole way science has been built up you know you've got to you've got to understand some things and then you build on that and then you build on that and then you build on that we're ways from getting there but I don't think there's any intrinsic limit to how far we can go humans are stupid but Humanity is smart yeah humans can be stupid but

01:27:54 humanity is smart very much and the gentleman right there when you mentioned the word philosophy and then before you wanted a definition of the word math I thought of the confrontation between the analytical philosophers and the Continental so the logical positivists victim Stein and Bertrand Russell thought math was the answer logic was the answer could this be a little bit of

01:28:24 what we're dancing around here yeah it was in everything I would say yeah I would say that and again Andrew and I've had had a number of conversations about this about what are some of the what any conversations like this what are some of the fundamental things and Russell had a famous quote about the the difference between mind and matter right and he said what is mind not matter what is matter never mind so that so the mind by the the mind-body

01:28:56 problem underlies all this the use of folk psychological language to describe scientific psychology our problems like this and I think that you know those go to the positive is the questions of the positive is about whether you could ultimately come up with a set of definitions and a set of you know non falsifiable statements and whether whether you could whether you could do that or not science hasn't while philosophy has moved past that and rejected that I think working scientists have never fully have never rejected the

01:29:29 logical positivist position as least as as a matter of practice I don't know if anyone has well you asked what do you think of as the alternative in your question to me was implicit that you know or is this just kind of skirting around logical positivism as opposed to what else should we perhaps be bringing in and considering that's the implicit question this artists seems

01:30:16 to us to have gap a besom between psychology and philosophy because it depends of where do we turn our attention and our focus of attention because this I think this gap is artificial because in order to I'll give you some example not the way that every experience requires our Wellness requires consciousness hit anyway for example Freud and Jung built his

01:30:48 here theory about the unconscious from the point of view not of unconscious from the point of view of conscious therefore he did you see some animal to write book about biology no man who is higher ontological creature right book about biology but I wished only to say that the UH not only the CMC Kathryn assertion not only our science excuse me sir I didn't roll so many books you don't ask questions I wish to say some

01:31:20 missing great point if contemporary physics is especially in physics because computation upon physics still jeans is the mass of jeans he tried to deviate this general intention of thought that the universe is not computed the universe is a mind because ever complete ever mathematics need of mathematicians what presenting itself mathematical equation without interpretation

01:31:50 okay this is not my question my question my team was to say something I think very important now we are aware that the whole science miss time temporality because see every all the computers so I

01:32:31 must respond or okay yes okay now that

01:33:25 happens that your comment one of the my comment about time thank you thank you thank you there one of the one of the people that I that we reference in the description of today's talk was Andre person who did supply that's a similar sort of analysis to making connection between the mathematics and and subjective experience so in that sense thank you very much and thank you everyone else here today for a really wonderful

01:34:03 [Applause]

01:35:29 [Music]