#51 [EN] Gautam Agarwal - "Como pode a investigação da Neurociência inspirar a construção de Inteligência Artificial mais avançada?"

Data: Wed, 06 Mar 2019 08:00:00 +0000  ·  Duração: 01:41:33

Clica em qualquer parte da transcrição para saltar para esse momento no vídeo.

José Maria Pimentel 00:00:00.00O convidado deste episódio é Gotham Agarwal, neurocientista doutorado pela Universidade da Califórnia, Berkeley. 00:00:24.32O Gotham está atualmente a fazer um pós-doutoramento no Laboratório de Neurociência dos Sistemas do Centro Champalimau em Lisboa. 00:00:32.12Aproveito para agradecer ao meu amigo Daniela de Lerba, que me o sugeriu, e na verdade já há algum tempo que queria trazer ao podcast alguém que fosse um exemplo destacado da vaga recente de investigadores estrangeiros a trabalhar em Portugal. 00:00:44.38Essa é uma tendência, aliás, que por si só dava um episódio, a crescente capacidade do nosso país em atrair alguns dos melhores cientistas em áreas de conhecimento com enorme potencial. 00:00:54.20A investigação do Gotham no Centro Champalimau tem passado sobretudo por usar jogos para estudar diferenças entre a maneira como seres humanos e algoritmos de inteligência artificial procuram soluções para os mesmos problemas. 00:01:06.16Este método tem o potencial de permitir, por um lado, compreender melhor o funcionamento do cérebro humano e, por outro, ao estudar os mecanismos do nosso cérebro que se mostram mais eficazes, obter daí inspiração para o desenvolvimento de inteligência artificial cada vez mais capaz. 00:01:19.90Conversámos, então, como não poderia deixar de ser, sobre estes temas, as forças e as fraquezas das versões atuais de inteligência artificial, mas também do cérebro humano, e a comparação entre estes dois modos diferentes de inteligência. 00:01:31.22E conversar sobre inteligência artificial é, na verdade, um regresso a um tema que já tinha abordado em episódios anteriores, em especial a conversa com a Linda Oliveira, que gravei logo ao quinto episódio. 00:01:41.22Durante esta longa conversa, falámos, então, sobre os desafios que subsistem na tentativa de criar inteligência artificial de nível humano, a chamada General AI, mas também, no sentido contrário, sobre as limitações da nossa própria cognição. 00:01:54.18A respeito do cérebro humano, que acabou por ser o que nos tomou mais tempo, conversámos sobre vários tópicos, desde o modo como aprendemos, ao papel do inconsciente, passando pelo mistério da consciência, a meditação, o recente interesse da ciência pelas substâncias psicadélicas, os dois hemisférios do nosso cérebro e, mesmo no final, sobre os motivos por que sonhamos. 00:02:13.64Antes de passarmos à conversa, uma nota rápida, não liga ao choramingar que se começa a ouvir perto do final, é que estava connosco também o filho do convidado, que tem um ano e cuja consciência não é ainda suficiente para ter noção do que é gravar um podcast, o que de resto só lhe fica bem. 00:02:31.96Gotam. 00:02:32.68Yes. 00:02:33.28Hã? 00:02:33.76Am I what? 00:02:34.32Like 99th percentile?
Gautam Agarwal 00:02:36.2696.3 percentile, yeah.
José Maria Pimentel 00:02:40.04Ok, fair enough. 00:02:41.44Ok, so as I said, instead of me delving into a specific topic right away, maybe it's more interesting to ask you about your aha moments in your research, either related to the brain or related to the way that we can infer from AI the way the brain works. 00:02:58.84So what was the, in these last years, especially in the years you've been linked to the foundation, what have been the major insights in your research and thinking?
Gautam Agarwal 00:03:09.30Well, I would say neuroscience for me is a very fascinating field because I think a lot of people who get into neuroscience are somehow inspired by their own personal experience because ultimately many of the things we study are phenomena that we experience on a daily basis. 00:03:30.64So like one kind of bad joke about, for example, people who study decision making, like me, is that we're very bad at making decisions. 00:03:40.92And so naturally, when there's something you're not very good at, you want to know how to do it properly. 00:03:46.42Exactly, yeah. 00:03:47.28And so in that sense, I feel like it differs from areas like quantum physics or astrophysics or let's say mathematics in that, in a way, we're in an experiment every moment of our lives.
José Maria Pimentel 00:04:02.82That's good, right? 00:04:03.78It means that you can complement the scientific method by sort of a more philosophical method in there you're looking inside and treating yourself as something which you can generalize from.
Gautam Agarwal 00:04:16.98Yeah, and I would say that that brings some of us into, let's say, a methodological quandary, which to kind of unpack that a little bit, science is very much about being able to set something up in a way that you can describe it very plainly to someone else and they can replicate it. 00:04:41.46So it tends to be very focused on the physical world and how to manipulate the physical world. 00:04:48.56Whereas when we talk about our experience, often it's very slippery. 00:04:53.22There can be a lot of biases in how we experience ourselves. 00:05:00.42So for example, to kind of give a touch on your previous question, what are some of the moments of insight? 00:05:07.86I would say there have been many little surprising moments sprinkled in the last four years I've been in Lisbon at the foundation because the foundation is a place where we can engage in many different types of experiments on ourselves and on our subject matters. 00:05:23.56And so one such one was called descriptive experience sampling. 00:05:28.60And this is a method that has been devised over many decades by a scientist in the US in Nevada. 00:05:37.66And the idea is that you have a little beeper that you wear on you and it goes off at random moments throughout the day. 00:05:45.72And at that moment, when the beeper goes off, you are supposed to write down in as much detail as possible the contents of your experience. 00:05:54.14And the reason that it's done like this is to avoid all of the preconceptions that come to the table. 00:06:00.54When I, as a neuroscientist, try to reflect on my way of thinking or way of experiencing the world, if I have a theory, that's going to contaminate what I have to report to you. 00:06:12.20And so what he has developed is a procedure where the beeper goes off. 00:06:18.76You write down, let's say, a paragraph about what is going on right now. 00:06:23.20So maybe I'm walking down the street and I have an English voice towards the left backside of my head that is thinking about the policeman who is looking at his watch. 00:06:37.86And that can be surprising because we usually don't have the moment to capture that as we're walking down the street. 00:06:45.54And over time with practice, with an interviewer, actually, the interviewer probes the description that you write. 00:06:52.98and picks out all the contradictions in what you've described. 00:06:56.58And over time, over the course of days, you get to the point where you can describe a tenth of a second in half a page. 00:07:04.32And so this is an example of an experiment that we did at the foundation with four or five of us and found that actually different people seem to have very different ways of experiencing the world as far as they're able to report them.
José Maria Pimentel 00:07:18.66Sorry to interrupt. 00:07:19.74It sort of links to personality traits, doesn't it?
Gautam Agarwal 00:07:23.74Yeah. 00:07:24.22So that's, I would say that's one question that we always want to come back to. 00:07:30.72Like, what is the meaning of all this? 00:07:32.42And when you are operating in such a relatively new space, like if you are a more verbal person, let's say, and your internal voice, let's say, speaks a few words and then pauses and then speaks a few more. 00:07:51.10What exactly does that relate to in terms of a personality, which can somewhat be a social construction, let's say? 00:07:59.78And this guy who developed the technique, he's collected data from thousands of people and he started to see what he thinks of as tropes in how groups of people think. 00:08:12.64So one example, it's, I would say it's not standardized, but one theory that he has based on his work is that people who have bulimia, this urge to purge, they're... 00:08:24.88They eat a lot and then they throw up. 00:08:26.44Exactly. 00:08:27.68It seems like that might actually be common or somehow linked to people who think in scattered thoughts. 00:08:37.74So a part of an idea starts at the back of their head and then moves to the front right of their head. 00:08:43.62And the bulimia, the purging moment, it allows their thoughts to coalesce. 00:08:49.70Very controversial, unclear what it means, but this is an idea of how an internal experience might tie it to something that we can all observe.
José Maria Pimentel 00:08:59.94That's great. 00:09:00.76That's really interesting. 00:09:02.28I mean, because you are connecting to completely different things and unexpected. 00:09:07.88And I mean, there's a lot of way to go there, to go to until you reach sort of a theory that will be able to explain that. 00:09:14.96And I asked about personality traits because it's a topic which interests me and I actually have an episode on that. 00:09:22.54And I always found it interesting about the connection between neurological differences, if you will, and which are like, which occur before or below in this construct. 00:09:39.18And then the end result, which is different personalities. 00:09:41.92Which are not, by the way, the only, this is sort of conceptual, but you can also talk about different values or different cognitions or different. 00:09:51.46So it's not all about personalities that have to do with this. 00:09:54.86I was thinking about that actually too, regarding the comparison of our brain, so the human brain and different types of AI. 00:10:04.62And there's always the assumption there that there's one thing called the human brain, which is sort of standardized, meaning it's the same between people, which is obviously not the case. 00:10:17.74And I was thinking whether it is fair to say that we should pay more attention, not only to the comparison between AI and our brain, but also between comparing AI with knowing that we have different types of different ways of thinking and different ways of building models of the world or different kind of different operating systems among ourselves.
Gautam Agarwal 00:10:42.04Yes, I think that's a beautiful point. 00:10:44.90And it's an issue that's at the core of neuroscience in some ways, because I think many scientific disciplines are about finding unifying principles. 00:10:57.90So, for example, you take physics, like the laws of gas. 00:11:02.38So you have, let's say, 10 to the 23rd molecules bouncing around this room right now. 00:11:10.32And we can model what's going on with just a few variables that have to do with the pressure, the temperature, the volume of the room. 00:11:19.58And in neuroscience, it's unclear how, to what degree can we make these simplifications, both in terms of the architecture of the brain, in terms of the cells that create the brain, and in terms of the behavioral outcomes of the brain. 00:11:36.00So as you're saying, can I think of you and me as roughly the same or roughly very different from each other? 00:11:44.16Economics would be an example that makes more of the assumption, as far as I understand, of a general commonality of behavior.
José Maria Pimentel 00:11:52.16Which is a big problem for economics, because you can never find the general truth.
Gautam Agarwal 00:11:55.66Yeah, and I would say at the same time, to me, that's where you almost have to start. 00:12:03.58You have to establish what's common before you can speak meaningfully about what's different. 00:12:10.66And luckily, now that we are borrowing more tools from mathematics, we can think about variability in more precise ways than in the past. 00:12:22.16So, for example, you mentioned personality for the last half century, if not for actually many centuries. 00:12:29.98If you look at even the Hindu tradition, there's ideas of different types of people. 00:12:36.76And so these ideas of types have come for a long time. 00:12:40.80And I would say what's kind of unique about this moment in, for example, neuroscience, is that we can tie this kind of verbal, these ideas that we have in verbal senses to more mathematically grounded ways of looking at diversity. 00:12:57.94And so, for example, there's many studies, including the work that I'm doing and many others around the world, of trying to collect data from tens of thousands of people, and in order to make precise statements that come from mathematics in terms of, okay, for example, types. 00:13:17.24There's this idea of the big five of personality. 00:13:21.64And these ideas, I imagine them to actually have quite a lot of truth to them, because a lot of people have thought very deeply about the origins. 00:13:30.30And you can do various tests to kind of validate them in terms of statistics. 00:13:36.86If I take the test today and I take it a few months for now, do I get roughly the same thing? 00:13:42.06But now we need to imagine the map of people's minds. 00:13:49.70And you can think of that map as a space, and each of our minds can be thought of as a point in that space. 00:13:57.52And now the whole set of points is, let's call it a cloud of some sort. 00:14:03.06And now the question is, is the cloud kind of a nice oval or like a sphere? 00:14:09.12Or does it have little branches where you have the autistic community in one little edge? 00:14:15.54Like clusters. 00:14:16.42Yeah, that's very far from everything else. 00:14:19.06Another kind of challenge in this is that we're very used to thinking of these distributions in low dimensions. 00:14:26.02So like a line or a plane. 00:14:30.42And actually, when you start thinking of distributions in high dimensions, very weird stuff starts to happen. 00:14:36.54For example, the points, you imagine a cloud in 2D, most of the points will be somewhere in the middle of the cloud. 00:14:45.70Now you go to 10 dimensions, let's say, how fast am I talking? 00:14:52.10How verbal am I? 00:14:54.04How mathematical am I? 00:14:55.50Each of these being a dimension. 00:14:57.26Now all of a sudden, all of a sudden, all the points are on the edge of the sphere, meaning that we're all outliers if we're in high dimension.
José Maria Pimentel 00:15:03.54That's very interesting. 00:15:04.56And it makes sense to some extent. 00:15:07.06Talking about this diversity, I think it may be useful to make it more concrete. 00:15:13.52So let's take, I mean, we could take memory or we can take perception. 00:15:17.96But for instance, learning, we tend to assume that the human brain is sort of built for learning in a way. 00:15:24.86And it's actually very interesting to compare that way in the way that machine learning, for instance, works. 00:15:30.68Machine learning being a specific method of AI. 00:15:34.00So like the combination of the two words. 00:15:36.04But it might not be the case, or is it? 00:15:38.14So do we human beings tend to learn the same way throughout our lives, and especially since we are born? 00:15:45.92Or are there like different softwares, if you will?
Gautam Agarwal 00:15:49.00Yeah, I wish I really knew the answer to that. 00:15:54.10But so you're raising issues of diversity that exist across people. 00:16:01.10So, you know, just in passing, you might have heard, I'm a very visual person, or I'm a very auditory person.
José Maria Pimentel 00:16:10.88I was alluding to that implicitly.
Gautam Agarwal 00:16:13.10Yes, and then I think another point that you're making is that in different contexts, we may learn in very different ways. 00:16:20.40So, for example, if you're sitting in a classroom, and you're in a geometry class, or literature class, or art class, or you're playing football with your friends, the set of, there's so much that's very different about each of these environments. 00:16:38.10So, one attempt at formalizing this in AI and in neuroscience is a topic called reinforcement learning. 00:16:47.34And the idea is that you can summarize all of the ingredients that contribute to learning in a relatively small set of categories. 00:16:57.48One of them is the states of the environment. 00:17:01.82Let's say, which part of the football field am I on right now? 00:17:05.60Another one is the actions of the environment. 00:17:09.36I can move to the left, to the right, front, or back. 00:17:12.84Another is the rewards. 00:17:15.50I get nothing for everything I do, almost, and I get one point for making a goal. 00:17:23.38And just with these relatively small set of ingredients, you can derive agents that are able to perform these tasks. 00:17:34.50And so, the problem is that these ideas, I mean, they're evolving very fast, and they're able to do things that have been people a few years ago would think are impossible for AIs to do. 00:17:48.34It's still a relatively very simple framework. 00:17:54.20And so, how does exactly this apply to a student in the classroom or you on the football field? 00:18:00.84There's a kind of gap between how we represent formally the knowledge that we provide to an AI that might help us understand your question of how learning is taking place and how you do it in everyday life. 00:18:14.44And so, at the moment, one approach in neuroscience and psychology is to really simplify the situations that humans are embedded in so that we can derive general principles of their learning. 00:18:29.40And the wager is that if I understand learning with you in a room where you're pressing a button to indicate if option A is better than option B, and you do that a thousand times, I can gain insights about your learning style that will generalize to everyday life. 00:18:48.70And that is, at the moment, we can't confirm that just because we haven't pushed the limits of how these experiments connect to what I experience on a daily basis. 00:19:03.64But I can say that it's amazing how much insight you can get from these simple experiments as far as their implications for everyday life.
José Maria Pimentel 00:19:14.40So, the limitation, I guess, there would be that you, I mean, we as adults, we are never learning on a blank slate, right? 00:19:22.98Which means that if I come here and you try to, and you put me in front of a TV or a computer screen, trying to get a learning process going on, I'm not coming with a blank slate. 00:19:34.60So, I have a life of past experiences and past learning that will be the building blocks of that learning process, right? 00:19:43.60So, I guess that would make it difficult to sort of completely separate the learning process that's occurring there and the cumulative learning process of your life.
Gautam Agarwal 00:19:53.22Yes, and in a way, I would say the goal right now of a lot of neuroscience work is to be able to make that separation. 00:20:02.48Can I create a task where your history before entering the room doesn't contaminate the results? 00:20:09.90As you're pointing out, there might be something very fundamentally important about the history that you carry out to that moment. 00:20:16.28But as a starting point, if we are to gain any sort of general principle about what a human being is like, let's say, on average, then that separation, that kind of compartmentalization of the human in the white box is a very important step. 00:20:37.44Now, at the same time, I would say the movement towards citizen science that I and various people are engaged in is to say, okay, do we really need to make that assumption of separating the person's history from their experience in an experiment? 00:20:57.04So, for example, there's studies where you look at a person as they behave in a laboratory environment, and you compare that to data collected using cell phones. 00:21:09.14So, now you download this app, you play this game or kind of gamified experiment on the phone. 00:21:16.76Do you do roughly the same thing when you're on, let's say, the metro? 00:21:20.94There's all these voices. 00:21:22.42You're thinking about five other things. 00:21:24.68Or as compared to when you're in a room, there's silence, there's nothing going on, and the surprising conclusion is actually they're kind of similar. 00:21:34.78And so, I think the question that remains is, is it similar because we aren't looking in a very careful way, or is it similar because that's all that human beings are, is what you put in front of them really dictates their experience?
José Maria Pimentel 00:21:53.62Yeah, that's, that's, that's quite interesting. 00:21:55.42And you have the, like, in comparing the way we think with the way, or the way we learn, and the way recent programs of artificial intelligence do it, I think there are quite a lot of interesting insights you can get from there. 00:22:12.58There's one which was, it was, it was very interesting for me to find out because I don't think a lot of people are acquainted with that fact. 00:22:21.02Although it can be challenged in light of the caveat I was mentioning before, which is that our brain is, it's really very, very efficient in learning. 00:22:32.48Which means that when we talk about all the breakthroughs that AI in recent years has made, often it's not mentioned the fact that those breakthroughs happen because a lot, a lot, a lot of inputs were fed to that AI program. 00:22:51.74Whereas us, human beings are able to learn it very fast. 00:22:55.54The caveat, the caveat for this is that, is the one we're talking about, which is that whenever we learn, we are not actually learning from scratch, but are kind of adding one block to, to, to the thing. 00:23:07.62There was something which I found very interesting in the paper you sent me about these things. 00:23:13.38And they, what they claimed was that even these recent breakthroughs in AI, so machine learning, neural networks, I mean, you, you, you, actually, if you can explain it, you'll explain it better than me. 00:23:26.12So these recent developments are actually based on sort of trend finding, whereas the typical human brain functions more under a permanent modeling of the world, which is, takes less time to, to, to get to, maybe less accurate. 00:23:45.26So may have a less, lower accuracy potential. 00:23:49.10So it may well be, and it has been overcome by, by AI in games like Go and chess for a long time. 00:23:56.12But it's more flexible. 00:23:57.60So if things change a little bit, we are able to adapt, whereas AI has had a lot of, I mean, there are a lot of examples. 00:24:02.56I remember one about, which was like the, you know, the self-driving cars. 00:24:07.22If you, if you put a sticker on the traffic sign, like a very small one, it was enough for the traffic, for the car to misrep, misapprehend it and take it for, I don't know, a different sign that it was supposed to be. 00:24:17.70So a very small trick. 00:24:19.06Yeah. 00:24:19.30We'll be able to, to trick AI.
Gautam Agarwal 00:24:21.18In a way, like, I think your summary is, is very accurate in terms of what some of the limitations are. 00:24:27.64Human, the brain, whether human or any animals are, is very efficient. 00:24:34.22So the, a lot of the work that is coming out in AI, it's done on these processors that are operating billions of times a second. 00:24:46.28And they're taking in thousands of images to produce an answer. 00:24:50.32If we had to put a number on the rate at which our brain is, is operating. 00:24:56.24Well, first off, it seems to operate at all, all rates in, in some sense. 00:25:00.80It follows kind of a power law distribution, which is, you see in, in the ocean, you have big waves and then you have little waves and they're all happening at the same time. 00:25:10.04There's no like fixed clock as you have in the computer and somehow it, it leverages, um, information that's coming in at different rates, at different granularities and, um, synthesizes them into some sort of, uh, continuous behavior in, in a way that can be so efficient that you can learn from a single example. 00:25:30.36Um, and, and so that is, um, part of the reason the authors, for example, in this paper suggest is that, uh, we're not, uh, just gluing a bunch of patterns into some sort of, uh, mega pattern that knows how to siphon data in through a series of pipes to produce the right answer. 00:25:51.10But rather we, uh, we, uh, come up with, we distill from, uh, thousands of input signals. 00:25:58.06So imagine even your eye right now, you have, uh, little dots, uh, each of which is, um, let's say bright or dim or colored or flickering. 00:26:08.36Um, and somehow you have these thousands of bits of information coming in and you can, uh, hone in on, let's say an explanation of what's happening in front of you, which is there's this neuroscientist. 00:26:21.52He's, um, uh, going to say something about AI to me and it's going to somehow fit into my narrative. 00:26:27.72You have kind of a story that you're telling that, um, really distills, um, an essence of what's going on and that story allows you to take new experiences and, and make sense of them in, in very efficient and creative ways. 00:26:44.84Um, and I think that's one of the major gaps right now that, that people are trying to figure out. 00:26:49.52And the flip side is that sometimes that's really dangerous. 00:26:53.86Um, so, and like, for example, in politics, this comes up a lot where, um, you have, uh, one idea that you hear someone say online, for example, and our learning apparatus is so powerful that you might completely flip, um, a switch in your brain that makes you think about something in a certain way, just from one piece of, of, uh, sensory input that isn't grounded in the statistics of the world. 00:27:20.38Um, so it's both a very powerful and a very dangerous, uh, weapon.
José Maria Pimentel 00:27:25.38The way I look at it, um, is that I think there's an insight to take from evolution, uh, looking at this. 00:27:32.22So the weight and the size of our brain, uh, is an evolutionary mystery in that it's consumes a lot of energy and it's sort of mysterious how, how, the way we were selected to have large brains. 00:27:43.56Obviously, we see the benefit of it, but also, we also see that it takes a lot of energy. 00:27:48.44So it must have been really very beneficial.
Gautam Agarwal 00:27:50.92A lot compared to our other organs and very little in term, in comparison to a computer.
José Maria Pimentel 00:27:56.58No, no, sure, sure, sure. 00:27:57.52Yes, it's much less than a computer, but a lot of energy, it consumes a lot of energy in comparison with, with the other organs. 00:28:03.50But what it means is that evolution was really very strict in terms of this energy consumption, meaning that the brain would need to be, to have, would have need to have evolved in a very, very efficient way. 00:28:14.96So consuming a lot, that's why we have these, these heuristics, these, I don't know if you, if you, to what extent do you buy that vision of like the two brains? 00:28:22.98So the, the, the sort of unconscious, automatic heuristic based brain, which is the one at play in the example you gave. 00:28:29.98And then the more conscious one that, that tackles, um, in a conscious, but very energy consuming way, a specific problem. 00:28:38.20The first one, the heuristics one, there must be another name for it, but the, the, the one that's, that is at play when we are driving, for instance, so automatically doing something that we already know it's, that's made to consume very little energy to be very efficient. 00:28:51.90But obviously that comes sometimes at the expense of, of biases, of errors, um, and that's, that's where, uh, that's a problem that the AI doesn't have, right? 00:29:01.42So that there's clearly something that we, that we are, um, at a disadvantage vis-a-vis, uh, AI, regardless of our, of our capacity to be flexible, uh, our, our fast brain is fast, but it's not very accurate.
Gautam Agarwal 00:29:15.82Yeah, I would say one, one, uh, another kind of surprisingly elusive idea is, um, what is it to be accurate? 00:29:26.88Uh, and, and there, I guess, like if you are asking someone to guess a number between one and 10 and you're, you're thinking five and they say seven, that's more accurate than nine. 00:29:39.78That's a fairly well-defined case. 00:29:39.78Um, in the case of a, a self-driving car, um, is it more accurate to, uh, drive off the street if there's someone in front of you? 00:29:39.78Um, and, and maybe you have an intuition for what the, what the right answer is. 00:29:58.64But my point is that, um, in everyday life, the, the concept of accuracy, which is fundamental to machine learning, to training a computer properly, that itself becomes very fuzzy. 00:30:12.14And, and, and that, you know, even though I agree that AIs in, in certain contexts are far less biased than, than we are, ultimately any bias they have is completely dictated by the programmer that chooses what it is to be accurate. 00:30:28.16What is the architecture of, of this AI? 00:30:28.16Um, what are the, how does it evolve over time?
José Maria Pimentel 00:30:34.54Sure. 00:30:34.54Yeah, that's why it's, it works. 00:30:34.54It has so far worked better in more sort of constrained situations, playing games, uh, and, and task-based situations rather than in the more, I guess, in the more complex situations of, of, of everyday life. 00:30:48.84In which AI would need to be able to perform in order for us to, to, to get to a general AI state. 00:30:48.84So general AI versus narrow AI. 00:30:48.84So meaning general meaning human comparable. 00:30:59.34I'm asking this because I know that you, you, you work a lot on, on like cross insights between neuroscience and AI. 00:30:59.34So things we are able to learn about our brain and can use developing AI and things we are able to learn from observing, sort of applying the same methods to AI and, and, and get those insights to our brain. 00:31:18.84The limitation there, I guess, um, is that we know we, on the one hand, we know that our brain is still better at a lot of things like this, the fact that it builds models that are more flexible, the fact that it is able to imagine different realities. 00:31:32.78So not, not just the reality that happens here that is happening right now, but, but kind of, kind of an alternative reality. 00:31:32.78We are able to, I'm able to imagine what you're thinking, a lot of stuff that we are able to do. 00:31:43.58But then on the other hand, we know about these limitations. 00:31:43.58So these biases, the limitations in actually in, even in processing power of our brain, uh, is it, is it not a huge challenge to be able to reap the insights that, that may be useful for AI without bringing along the sort of a defects rather than characteristics of our brain?
Gautam Agarwal 00:32:03.08Um, so you're talking in the direction of AI to neuroscience?
José Maria Pimentel 00:32:06.54No, no, no, no. Neuroscience to AI. 00:32:06.54So studying, basically, we, we, so there's a fundamental question here, which is what's the way to build general AI? 00:32:15.34So one, one way is to get inspiration from the brain, which is the only intelligent thinking machine we know of. 00:32:15.34Uh, there's another way, which is going completely around that and building it from scratch based on, on different principles. 00:32:27.78But if you, if we follow the first one, so getting inspirations from the human brain, uh, I guess it will, it will be very tricky not to bring along with the good things of the human brain, the bad things, if you will. 00:32:39.22So the, the, the defects of the human brain, the, the shortcomings of the human brain.
Gautam Agarwal 00:32:42.76Yeah. 00:32:42.76For me, I guess one, one challenge in all of this is that we are trying to build something with, uh, where the principles of engineering are still not established and we're trying to get it to solve a problem that isn't very clear. 00:32:59.66So if you have, um, for example, you want to build a bridge, then you can take the, you have some sort of objective, which is to cross from the, you know, the North to the South of the river. 00:33:08.56It needs to be resilient according to, um, certain constraints. 00:33:08.56And then you can optimize for, for these physical constraints to develop a bridge. 00:33:22.72When you train AI on something relatively straightforward, like a game, you know what the objective is. 00:33:22.72Um, even maybe for something like a self-driving car, you can say, minimize the number of, of accidents. 00:33:22.72Um, but, but there in, when you start seeing some of the ambiguity of the, the high level principle, uh, creep in, and this is, I would say even more so the case for this concept of general AI, what exactly are you trying to optimize? 00:33:52.30It kind of goes back to the start of our conversation. 00:33:52.30Even if you think about your own experience, it's a very slippery concept. 00:33:52.30You know, who are you? 00:33:52.30What are you? 00:33:52.30Do you know what you are optimizing for? 00:33:52.30Yeah. 00:33:52.30Yeah. 00:33:52.30I certainly don't know what I'm, yeah. 00:33:52.30Um, and so I guess what, what I've been very interested in is in, in my own work is to try to come up with, um, environments that are not too simple because simple. 00:34:22.30environments are easy to solve and can be solved trivially. 00:34:22.30You don't really need intelligence to solve a very simple problem, but they're not too complex so that you get, you lose sight of what it is that you're trying to optimize. 00:34:35.30And in a way, you know, we talked about games like, let's say, uh, Atari games or go as being relatively simple from a mathematical standpoint, they're actually very complex. 00:34:35.30Yeah. 00:34:35.30Yeah. 00:34:35.30Sure. 00:34:35.30And so, um,
José Maria Pimentel 00:34:48.88I think I used the word constrained, maybe not, but that's what I meant. 00:34:48.88So simple in the sense that not being simple to play, but, but, uh, happening within a, uh, constrained set of rules. 00:34:48.88Yes. 00:34:48.88Rather than in comparison with, with our life, which is much more complex. 00:34:48.88Yes. 00:34:48.88Complex is an ambiguous word. 00:34:48.88And it's a very relative word. 00:34:48.88Yeah. 00:34:48.88Um, so, uh, a very kind of, um, important concept in, in, um, AI and the intersection between AI.
Gautam Agarwal 00:35:18.88And neuroscience right now is the curse of dimensionality. 00:35:18.88And, um, the, um, the idea there is that even if you start with a very constrained environment where, um, I can only, uh, have a controller, I'm playing an Atari game. 00:35:18.88I can move left, right, or jump. 00:35:18.88Um, okay. 00:35:18.88So that's just three actions. 00:35:18.88Big deal. 00:35:18.88But now every time step, every, let's say you're playing at, uh, uh, 10 times a second, you're thinking about what to do. 00:35:18.88Uh, that's, Three times three times three times three times three times three times three. 00:35:48.86And within a second, you have, uh, 5,000 different actions that, that you need, that you could possibly take. 00:35:48.86And so if your goal is to find the best string of actions, you're searching among 5,000 different possibilities at every moment in, in time. 00:35:48.86And that's just if you care about one second. 00:35:48.86Um, and so, uh, this is the, the challenge whereby these seemingly simple games, simple games, constrained games become immensely complex. 00:36:18.84Um, and so, uh, one thing that I've been using is games that are closer to, uh, the game tic-tac-toe. 00:36:18.84I don't know in Portuguese, uh, what it might be referred to as. 00:36:18.84Do you know this game? 00:36:35.64Um, it's called Jogo do Galo. 00:36:35.64Okay. 00:36:35.64The rooster game. 00:36:35.64Yeah. 00:36:35.64Yeah. 00:36:35.64So, um, Jogo do Galo is kind of Jogo do Galo. 00:36:35.64Galo. 00:36:35.64Galo. 00:36:35.64Galo. 00:36:35.64Yeah. 00:36:35.64Yeah. 00:36:35.64No, you pronounced it correctly. 00:36:35.64Okay. 00:36:35.64Cool. 00:36:35.64Yeah. 00:36:35.64So that is a game that is, um, much simpler than some of these Atari games. 00:36:53.64Um, and even there, you, you kind of get into, if you think of what's called kind of, um, a fitness landscape, there's this concept in evolution, for example, uh, which genes, uh, lead to the most fitness. 00:36:53.64So, and this, uh, image of a landscape, the idea is that every point in this landscape represents a different configuration. 00:37:19.52Uh, in one case, it might be, in the evolutionary case, it might be, you know, your genetic makeup. 00:37:19.52And then the high points of this landscape correspond to very good genetic makeups. 00:37:19.52Uh, and this is also the case in, um, in AI. 00:37:19.52If you're thinking about the possible actions that I can take, I can map out those actions on a landscape and the high points of those landscapes correspond to the best sequence of actions. 00:37:47.10So even a game like, uh, Jogo de Galo is one where, um, you have actually quite a, um, a large landscape and, and you have hills and valleys in that landscape. 00:38:00.34And, uh, your goal is to get to the highest point of that landscape, but you need to go through a valley sometimes before you can get to the peak. 00:38:08.42And that's something that can be a problem for some of the AI systems because they just follow the, the direction that increases their height. 00:38:18.10And so when, when, when we learn games like, uh, Jogo de Galo, we learned that, okay, we need to connect three in a row. 00:38:26.44We abstract away this concept so that it doesn't necessarily have to be the left vertical column. 00:38:33.18It can be diagonal, it can be going across and, and then all of a sudden we are connecting different disjoint separate points of this value landscape to the same concept.
José Maria Pimentel 00:38:45.82That's actually the way we, the way we, the way we, the way we are able to memorize so many things is by connecting and generalizing, uh, uh, a lot of observations into a general principle.
Gautam Agarwal 00:38:57.10Yes. 00:38:57.10And the, I think for me, one of the miracles is that, um, the, the way in which we glue these ideas together sometimes can be very removed from our naive mathematical ways of describing them. 00:39:13.36Um, so, um, this idea of a landscape, if I just put one X down one over to the right or two over to the right, it can mean something very different and put you in very different parts of the landscape. 00:39:24.80And somehow we're able to glue the right parts of the landscape together. 00:39:28.76Um, and, and so the idea that, that I've, um, I've been studying for the last few years is, um, now we also have a landscape as an experimenter of games. 00:39:40.48Um, so there's very simple games and very complex games. 00:39:43.76And I think we might agree on the limitations of going too much on either extreme. 00:39:49.06What is a good middle ground where I can both challenge your intellect, your intelligence and how you find creative solutions to hard problems, but model it mathematically, make it such that by looking at, uh, thousands of people, I can start finding general principles of, okay. 00:40:10.30This is a type of person that, um, forms this kind of rule in terms of figuring out how to play this game efficiently. 00:40:19.74Uh, or maybe there's groups of groups of people who don't form any rules at all. 00:40:24.38They use more of an instinct already in the lab. 00:40:27.42When of the people I've looked at, you see people who are very committed to this rule-based approach or vice versa. 00:40:34.08And, um, but until I get sufficient statistics, I, I can't really make claims about that.
José Maria Pimentel 00:40:40.48No, and it brings us back to what we were talking about before about personality. 00:40:43.42For instance, I, I, I see myself, maybe I'm wrong, but in comparing with the people I interact with, I see myself as typically someone who learns by general, establishing general principles, sort of a big picture sort of approach, if you will. 00:40:57.76And I think there are people who follow different approach, more of a low level approach, uh, in that they are, but they use instinct, if you will, or they, they, they collect more granular data rather than trying to understand like the, the overarching principle of, of the thing you're trying to, to memorize. 00:41:14.36So maybe people, not everyone behaves the same way when learning or when memorizing.
Gautam Agarwal 00:41:19.42And there's the idea from an evolutionary perspective, the so-called term is neurodiversity. 00:41:26.44Yeah. 00:41:26.78That actually it's good that you learn in a very different way than someone else, because that'll allow you to pick up certain things that they want and vice versa.
José Maria Pimentel 00:41:36.26No, that's, and in real life, it's clearly, clearly true. 00:41:38.74So every, every approach is by definition has limitations. 00:41:42.90So it's, it's limited by definition, which means that if I think this way and you think in a different way, we'll be able to complement each other because you're looking at something like your approach. 00:41:51.66We'll be able to get, uh, uh, insights that my approach won't be able to and vice versa.
Gautam Agarwal 00:41:57.08And I think kind of the, the opportunity that we have in neuroscience right now in the kind of intersection of AI neuroscience and psychology is to take this intuition that this diversity is somehow useful. 00:42:11.64And these verbal descriptions that I think in a more broad way, and there's people who think in a more granular way, can we, uh, write, uh, put some numbers to this? 00:42:22.32Um, can I actually map, can I visualize what your big picture way of learning looks like? 00:42:29.08Because, um, another kind of aspect of diversity is that we gain very different insights when it comes from a verbal description and a visual or a sensory description. 00:42:39.94So right now we're engaged in verbal description. 00:42:42.84What happens if, uh, on a, you know, virtual reality map, I can project a shape that represents the style in which you learn and compare that to that of other people. 00:42:53.38I think we'll gain different kinds of insight into this kind of conversation.
José Maria Pimentel 00:42:57.38Yeah. 00:42:57.62And the way we, the way we memorize too, for instance, I've, I've, I've, I heard something on one of the podcast episodes we were talking about before the conversation, um, that I found very interesting that the, the, the, this guy was, was explaining that there are basically two ways in which we memorize. 00:43:14.58I'm not sure if, if this is, um, if this was his opinion or whether this is, uh, something which is clear evidence. 00:43:22.28So the standard way is for us to sort of the long-term memory, it, it, it happens by, by synapses, uh, between the neurons so that, that are flashed. 00:43:32.14And then when we flash them again, we remember things. 00:43:34.84And there's, then there's another way for the working memory or short-term memory when you have to memorize the phone number, for instance, in which it happens, it happens in a different way, different way. 00:43:43.80It happens at the level of, of neurons, uh, isolated, I think. 00:43:47.56So rather than, it has more to do with, I'm stepping to uncomfortable terrain here. 00:43:52.90Go for it. 00:43:53.80For my, for my knowledge, but it happens at the level of neurons rather than at the level of, of synapses. 00:43:59.58And it's by definition, short-term. 00:44:01.98So if, if I ask you to memorize a 10 digit number, you're able to do it. 00:44:05.62But then if I ask you again for it, uh, tomorrow, you won't remember it or, or, or would I? 00:44:11.24Because, because you are working with your short-term memory that's, that's sort of short-lived. 00:44:17.58Yes. 00:44:18.26Is, is this correct or?
Gautam Agarwal 00:44:19.80Yeah. 00:44:20.12Um, and so the, the concept of short-term and long-term memory, I think is a very useful way of, of categorizing different styles of learning. 00:44:29.28Now it turns out that there's actually, learning seems to also be one of these things that's happening on every timescale. 00:44:36.28So, um, if you look at, um, cellular processes, you have, um, activity at the synapse that starts, let's say, with an action potential, which is a one millisecond event. 00:44:49.96And that causes the release of neurotransmitters over the course of tens of milliseconds. 00:44:55.18And that invokes, uh, calcium responses over hundreds of milliseconds, which invokes changes of configurations of proteins and chemical tagging over the course of seconds that can lead to gene expression over the course of tens of seconds or minutes, et cetera, et cetera, et cetera, until you end up with years. 00:45:15.06Yeah. 00:45:15.58So in, I would say the terms short and long-term are colloquially useful and they invoke a multi-scale biological process at all timescales and possibly is another example of the diversity that we were talking about in the expression as expressed in learning, where if you really dissect, um, the different phenomena involved in learning, you might end up with way more than, uh, two categories that fall on one dimension. 00:45:44.74Um, as we said before, um, as we said before, is muscle memory on the football field, the same as, um, uh, cognitive memory as established when thinking rationally, for example, even if they're at the same time scale.
José Maria Pimentel 00:45:58.30Sure. 00:45:58.30No. And ultimately it's, it's really, hugely complex. 00:45:58.30So the, for instance, in the brain, as you were, as you were alluding to there, a lot of things happening at the same time, but it's, I guess it's not easy or, or, or rather it's, it's almost impossible to, to separate everything and be able to pinpoint. 00:46:14.42Pinpoint what's, what's, what's the origin of, of, of, of these memories, for instance, or, or, or, or something else. 00:46:14.42Um, but memory for me is actually quite, it's, it's quite fascinating. 00:46:14.42I've, I've thought a lot of times about that because I, I'm always puzzled by the way. 00:46:28.42There are some things we, there are some things we forget within months and other things that stick to our, to our memory. 00:46:28.42It's, I mean, for me, it's, it's a big mystery. 00:46:28.42I have one, actually, I have one theory that it would be cool to have your, your take on. 00:46:28.42I think it has, well, this is sort of obvious. 00:46:28.42It has to do with, with the importance we give to it. 00:46:50.42But it's more the unconscious, unconscious importance we give to, to something. 00:46:50.42Say I'm reading a book. 00:46:50.42Say I could read the book into, into parallel realities. 00:46:50.42One of them I'll be reading the book at the beach, for instance, or sort of relaxed, uh, non-committed way. 00:46:50.42Uh, and in the other reality, I'll be reading the book in, in preparation for this conversation. 00:47:11.42What happens with me, and I have a lot of data points for this, is that if I, in this second situation, uh, I'll remember, uh, much more a year from, from that day. 00:47:11.42It happened, for instance, I was looking at the note I had taken for, you know, the other, the other conversation I, I, I have, um, related to AI for the podcast. 00:47:30.74We'd have, which happened in, I think, October, 2017. 00:47:30.74So more than a year ago. 00:47:30.74And I remember a lot of that stuff because I was, I was reading things that were useful for me in the, in the short run. 00:47:30.74Uh, and, and sort of my unconscious was committed to that, that learning process. 00:47:30.74But, but I'm not completely sure of this insight. 00:47:30.74Does this make sense?
Gautam Agarwal 00:47:51.66Yeah, it does. 00:47:51.66Um, I, one thing I'm curious about is, um, what is the advantage of reading on the beach? 00:47:51.66Like what is, um, what might be better in that scenario for you? 00:47:51.66Like, no, no, no, it's worse. 00:48:06.34Yeah, I see what's worse about it, but are there any advantages, um, to reading on the beach?
José Maria Pimentel 00:48:12.80It's more relaxing. 00:48:12.80So I think that the second scenario yields ultimately longer term memories because you're, you are wasting your waste. 00:48:12.80You're using more resources in that scenario than in the first one. 00:48:12.80So reading in the relaxed way is cool, but you're not, you're not taking as much from it as you, you take the relax, the relaxation from it, but you're not, you do not take the, the, the, the insight. 00:48:35.58It's sort of the, it's analogous to that principle, which became known from, um, Malcolm Gladwell book. 00:48:35.58It's called, what is it? 00:48:35.58Deliberate practice. 00:48:35.58You know, that thing about the 10,000 hour rule. 00:48:35.58It's, it's, I'm pretty sure it's much simplified, but it has to do like the deliberate word that comes before practice. 00:48:35.58It's quite important there because it means that you are, well, to summarize it, it means that you are, you, you need supposedly 10,000 hours to be proficient at something, but that doesn't happen. 00:49:05.58And it happens just by you doing that. 00:49:05.58It happens by you doing that in a deliberate way. 00:49:05.58So you are completely committed to, to, to, to doing that. 00:49:05.58You're, you're, you're, you're, the, the full resources or the full potential of your brain is committed in doing that rather than just reading for fun or, or, or playing guitar in a, in an uncommitted way. 00:49:25.76Uh, but at, so this is more of a psychological insights. 00:49:25.76Uh, but I don't know if this has, if this has been sort of proved that in a more neurological, uh, um, way of analysis. 00:49:25.76So analyzing the brain rather than like the end process.
Gautam Agarwal 00:49:42.40The kind of, um, I guess the closest, uh, phenomenon that I can think of is that of attention and that people have studied, um, scientifically, both psychologically and in neuroscience. 00:49:42.40And, um, there seems to be kind of varying effects of paying attention to something. 00:50:00.30One is that let's say I'm paying attention, um, to my visual experience more that can, uh, raise the activity of certain neurons. 00:50:00.30It can make them more, uh, respond in higher fidelity, which means that if I show the same object 10 times, the neuron will be more consistent in how it responds to this object. 00:50:26.20Each of the 10 times when you're paying attention. 00:50:26.20And also, as you're saying, um, it can, uh, increase the, the learning associated with it. 00:50:26.20The idea is, I think also related to that of limited resources. 00:50:26.20So at the end of the day, you have all of this information coming in. 00:50:26.20How do you choose which of it is relevant? 00:50:26.20Yeah. 00:50:26.20And I guess coming to yet another mystery is why does it have to have to be conscious? 00:50:56.18Like why, why couldn't it be something that your body does automatically? 00:50:56.18Why do you need to, why does it seem to be tied to your conscious awareness of what's going on?
José Maria Pimentel 00:51:07.50So you want to talk about consciousness? 00:51:07.50Is that what you mean? 00:51:09.62Yeah. 00:51:09.62Yeah. 00:51:09.62I've been itching for that conversation for some time. 00:51:13.54Okay. 00:51:13.54So, okay. 00:51:13.54So I'll make the introduction. 00:51:13.54I think it's a great segue for that. 00:51:13.54It's a topic I love, uh, actually, and because it relates to, it may be a limitation of AI. 00:51:13.54We never know because it's very hard to, to, to, to define consciousness. 00:51:26.16For me, it's, it's natural, um, to recognize it. 00:51:26.16Although being an emergent process, it's something that's, for instance, I've heard some people from, from the AI world claiming that the expression was that consciousness will be explained away. 00:51:26.16Meaning that I think the way I interpret it is that it will, at the end of the game, it will disappear basically because it's not, it's something very diffuse that that's not actually there. 00:51:26.16It's just the effects of, of, of different things. 00:51:56.14For instance, um, um, Dan Dennett, the philosopher, I love his, a lot of his work and the, he had, the way he defines it, uh, is that consciousness in the, in the way of comprehension. 00:52:08.14So really the way, the fact that we understand what we are doing here and we tend to look at some sort, some, at least the AI systems we have so far as not understanding what they are doing. 00:52:08.14So in doing it in a, in a mechanical way, but not having awareness. 00:52:08.14He says that comprehension slash, uh, ultimately consciousness comes after competence. 00:52:08.14So it's, it's an effect of competence. 00:52:08.14So if you have, in our evolution, we gain competence. 00:52:08.14And this is kind of the byproduct of competence. 00:52:08.14And nothing, nothing more than that.
Gautam Agarwal 00:52:38.12Yeah. 00:52:38.12I suppose, um, that's probably one thing that I find very interesting about it as a topic is that, you know, we all have some sort of, um, we, a lot of us identify with our own conscious experiences. 00:52:38.12And as, uh, as a community, a scientific community, there's no consensus on, on what it is. 00:53:00.84And so it's both incredibly accessible and universal and, uh, incredibly elusive at the same time. 00:53:00.84Um, so from the neuroscience perspective, I would say a lot of what we do, um, comes from the very physicalist stance. 00:53:00.84Um, generally, I think many, uh, neuroscience as, uh, at a community level, we operate on the principle that the brain is a physical machine, let's call it. 00:53:27.86And that, um, and that, um, any, um, any, any expression that we are capable of or experience is ultimately, uh, a side effect or, uh, an outcome of that, the workings of that physical machine. 00:53:27.86Um, I would say that, you know, it is a good working premise in terms of how we conduct our experiments. 00:53:51.86To me, to me, it doesn't get any closer to answering the mystery that has confronted, uh, our cultures for thousands of years. 00:54:00.44And in fact, has been talked about in very sophisticated and diverse ways in many different communities, let's say, including, um, religious practices, um, and including art. 00:54:14.76I would say there's certain levels of insight that, um, um, I would say they, they have to offer about that process that come from very different places. 00:54:24.42Philosophy is another example. 00:54:25.96So if you, um, talk to, um, or some philosophers, uh, categorize the different stances on what consciousness is. 00:54:36.16Um, and most scientists, um, fall into one corner of that space of possibilities, um, let's call it the physicalist, monist, reductionist corner. 00:54:49.10Um, then there's the, the mental monists who think of experience as being the primary phenomenon. 00:54:57.76So, you know, many of us grow up in this, um, working premise that, okay, this is what I touch, uh, this is the physical world. 00:55:05.40We can create devices. 00:55:06.56We clearly have some sort of, um, control. 00:55:09.72If something happens to me in the physical world, I will pay for it. 00:55:13.60Um, you know, all of these things lead us to be assumed that the, the physics is first. 00:55:19.56Um, if, uh, if you're coming from a more Buddhist perspective, it would be actually the experiences that I have right now, which are a very mental thing are what's first. 00:55:31.16And what I find really fascinating about that approach is that actually that can be seen as a science too. 00:55:38.00There's, um, many, uh, terms of different flavors of attention or different flavors of emotion. 00:55:45.62There might be many dozens of terms that, um, have been created by that community that can really take this seemingly nebulous construct of experience and really partition it and also operate on it in systematic ways. 00:56:01.88If I'm in this state of, um, desire, how can I go from this state of desire to one of, um, acceptance? 00:56:10.54You know, and there's, uh, in, in terms of mental experience, they might be quite far removed from each other, but there's a certain discourse around a systematic way of going through that space. 00:56:22.34And, and, and so one kind of more recent development that I'm very excited about is the conversations that are going on between these different communities. 00:56:31.64So even the Dalai Lama has gotten involved in conversations with neuroscientists where there's camps, neuroscientists going to teach monks about ideas in neuroscience and vice versa. 00:56:43.14And, uh, the hope is that there might be some, uh, cross fertilization in, in that.
José Maria Pimentel 00:56:48.78Yeah. 00:56:49.04And there are a lot of insights that have been gained from, from Buddhism and meditation, right? 00:56:53.64So from, from being able to sort of observe your own consciousness, um, removed from it in a sense.
Gautam Agarwal 00:57:01.82Absolutely. 00:57:02.24And, uh, you know, I think one thing that I find to be very missing in the physicalist approach is that it can, um, rob our lives of meaning in some way there, there can be, you know, if everything is, um, a cause and effect chain of atoms bouncing around and influencing other atoms, um, that can be, uh, somewhat of an alienating experience. 00:57:29.40And, and I think spiritual traditions, um, they can offer psychological mechanisms that allow us to, uh, operate in, in the world, uh, in a way that gives us more agency over emotions. 00:57:45.60And, and, and so in that sense, I think in a functional sense, it's very useful and it actually gets at some of the, um, issues that neuroscientists are very interested in, uh, for example, emotion. 00:57:57.48So in that sense, even though these ideas are coming from a very different place, they might have these systematic, um, insights that we can borrow.
José Maria Pimentel 00:58:05.82So how do you view emotion, by the way, what's your, what's your, what's your, this, this may seem like a basic question, but what's your, uh, because it's, it's a word that's, that's used for so many things. 00:58:15.88What's your, what's your take on it in our, in human cognition?
Gautam Agarwal 00:58:20.06So right now, the, the strongest association in neuroscience, uh, with emotion is that of neuromodulators. 00:58:29.02So these chemicals that include, for example, serotonin and dopamine and adrenaline and acetylcholine, they seem to have, um, uh, centers in the base of our brains composed of these neurons that sprout and connect throughout the brain. 00:58:50.42And at certain moments, let's say moments of surprise or moments of reward, they, um, become active and release. 00:58:58.40They kind of bathe, they seem to bathe the brain in a certain neuromodulator in ways that seems to correspond to emotion. 00:59:06.14One of the strongest connections that has been established is between dopamine and, um, and the experience of reward. 00:59:14.52Um, and so it seems like, um, there's connected, sorry to interrupt, it's connected to, to the
José Maria Pimentel 00:59:19.98learning process we were talking about, right? 00:59:21.48Exactly. 00:59:21.98As a reinforcement mechanism of, of the learning.
Gautam Agarwal 00:59:25.06That's right. 00:59:26.04Um, and so, uh, there's, uh, it seems like there's been a lot of progress. 00:59:31.68Yes, I would say the, the story is probably, uh, much more complex as any story is. 00:59:38.46And, um, so for example, is dopamine always linked to these, these particular emotions? 00:59:45.90What if it is a war, a reward that you're ashamed of? 00:59:50.28For example, you know, there's all these, um, these contingent emotions that like, for example, certain writers might be very careful in describing and, um, I would say the current map that we have in neuroscience, uh, that, uh, connecting neuromodulators to emotions doesn't quite capture.
José Maria Pimentel 01:00:09.80This came up to my mind now because we were talking about the Buddhism and that leads to meditation. 01:00:14.86What's your, your view on, on being able to sort of alter the way your, your brain behaves, uh, from the outside. 01:00:23.34So we talk about thinking, uh, intelligence, ultimately consciousness as emergent processes that emerge from your brain, emerge from your brain activity. 01:00:33.70Uh, but one thing Buddhism and meditation tries to do is to also reverse that. 01:00:38.94So sort of by meditating, for instance, you're able to alter the way your brain works and this can work to reduce anxiety, for instance, is one of the applications as, as a neuroscientist, how do you view that? 01:00:51.70Yeah.
Gautam Agarwal 01:00:52.08Well, we, um, a bunch of my colleagues and I, we, um, started a meditation working group. 01:00:58.64We met for about a year or so, um, invited a teacher and went through the, the, um, experience, um, focused on mindfulness and in fact, inspired by a cognitive scientist, Mark Williams, I believe, um, who gave a talk at our, um, center. 01:01:16.08And, um, I would say, you know, like any community, well, science, science is definitely a community where, um, we, uh, we face, uh, emotional challenges on a daily basis in the sense that we are working towards a goal, which might not produce, uh, viable results for years. 01:01:35.52And so it might be no surprise that there's actually a lot of, um, mental health issues in the scientific community. 01:01:43.92Um, graduate students, uh, experience a lot of anxiety and depression. 01:01:47.96And, and, and so there's a little bit of an irony in that the field that studies the brain, um, a lot of its practitioners have these, these problems in a sense. 01:01:59.20And, and I, I would say I'm very much someone who's had to cope with these experiences in the uncertainty of the scientific enterprise.
José Maria Pimentel 01:02:07.54Um, and, and so I would say my field, every field, which, which has uncertainty, uh, is related with, with, with challenges, uh, with emotional challenges. 01:02:18.14Uh, so that's why every, every creative field, for instance, if you are writers, painters, they're, they, they're all people who typically are, are, have to, to fight cases of depression or, or, or anxiety because you are, you live in an uncertain world, uh, being it's related to artistic creativity or to scientific research in your case.
Gautam Agarwal 01:02:39.72Yes. 01:02:40.06And, uh, I think uncertainty, um, moments of, of challenge where it's not uncertain, it's like definitely bad, for example. 01:02:48.82Um, these are all cases where, um, I'm certainly at a personal level, very curious about what meditation has to offer. 01:02:57.04And in a way it, um, from a neuroscientific perspective, it's fascinating because at least the implication for mindfulness is that just by paying attention, not trying to do anything about it, it might actually help you achieve, uh, a certain goal in terms of your emotions are as, as far as they're concerned.
José Maria Pimentel 01:03:18.46Yeah. 01:03:18.60It's, it's very puzzling to me. 01:03:20.22Uh, I have tried meditation only a few times, uh, but I have to say not, not, um, not very deliberately to, to pick the word I was using before. 01:03:29.18Uh, so I, I wasn't able to gain the, um, those insights yet, but I've, I've been very curious about it. 01:03:34.98The, the, the, and meditation, it's only one way of accessing, um, consciousness. 01:03:39.76We talked before the conversation about, um, something which I also found very, find very interesting, which is the case of psychedelics. 01:03:46.10So being able to access your consciousness in a way, maybe in a quicker way, in a, in a, I would wager not in a, in such a profound way, maybe I'm not correct because meditation occurs like for you to reap benefits of meditation. 01:03:58.80It's, it takes a lot of time. 01:04:00.10It takes really long time and the benefits, uh, get higher and higher, uh, the more you do it, uh, but they, they only start happening. 01:04:09.00I think in, uh, within weeks or even months, whereas psychedelics are, are, are an immediate thing. 01:04:16.10And first I was puzzled because I, I, I saw a lot of unsuspecting people talking about that. 01:04:23.20So philosophers, scientists, I mean, Sam Harris, for instance, which, who is a guy who, whose podcast I listened to a lot and even a very stern guy in a sense. 01:04:33.50So in the sense that he's not very outspoken. 01:04:35.96So, and, and he was talking about that and gaining a lot of insights from, from, from that, uh, and especially psychedelics being one thing that is typically, at least in Portugal frowned upon. 01:04:47.02So it's, it's, it's, it's basically, it's treated in the same way for, for, for the common folk, I guess it's treated as if it were strong drugs or even, or, or, or, or, or not so strong drugs, but it's treated in the same way. 01:04:59.24Whereas in this field, it's seen as a way of, of getting a lot of insights from your consciousness.
Gautam Agarwal 01:05:04.72I think it's a very profound vehicle. 01:05:08.80Um, there's, uh, it's one of the, the cases where a specific molecule, which is a physical entity can induce a, a very profound, a very intimate and a very universal experience, allowing two very different people from two very different backgrounds to, um, develop a language about very internal insights.
José Maria Pimentel 01:05:42.52And it happens by, by, as I get it, it happens by sort of turning off parts of the brain, right?
Gautam Agarwal 01:05:48.62Yeah, it's, well, it's, it's hard to, it's really hard to say, um, what exactly is going on. 01:05:55.40There are, it seems like we're in a time period where, um, after about a 40 year drought, um, there is, there are more scientists around the world working on the exact mechanisms and the implications of, uh, psychedelics. 01:06:12.54And, um, the current studies suggest that, uh, at least for certain psychedelics, um, parts of the brains, the, their activity is greatly reduced. 01:06:22.50Now, what that means is, is quite unclear. 01:06:25.98Yeah.
José Maria Pimentel 01:06:26.14The way I see it, but well, this, I'm obviously a layman, so, so this might be completely wrong, but the way I see some, some parallels with dreaming because in the way I interpret my, my dreams, uh, and I was actually talking with, with friends about this today is that it's as if a part of my brain is turned off. 01:06:45.20So for the, the part of the brain, which checks the, the plausibility of things that at least is turned off. 01:06:51.58So very strange things happens in, happen in my dreams that are not plausible. 01:06:56.56And in real life, uh, I would find that, that, that they were not, uh, possible right away. 01:07:02.32And in dreams, you, you, you can go on having a dream that it's unplausible, uh, until you wake up. 01:07:08.30Uh, and I would guess that happens because a part of your brain is, is turned off, but maybe it's a bit more complex than that.
Gautam Agarwal 01:07:15.34Yeah. 01:07:15.60And I think part of the complexity arises from what does it mean for something to be off? 01:07:21.94So when we think of something being off, often it's in terms of a switch being off. 01:07:26.68That means there's no electricity going. 01:07:29.34It tends to be kind of a neutral state. 01:07:32.26In, in, in neuroscience, um, off actually is, uh, is a much more, uh, active concept in a sense, because, uh, something is, something that's on is only meaningful in far as it turns off sometimes. 01:07:48.54And so, um, those states of offness can actually convey, um, a lot of information.
José Maria Pimentel 01:07:55.10As much as, as being on. 01:07:56.68Yes. 01:07:57.22That's what I mean. 01:07:57.80Yeah. 01:07:58.02It makes sense. 01:07:59.66It makes complete sense. 01:08:00.62Um, let me, uh, try and close this topic of consciousness. 01:08:04.54There, there's one thing that there's one way, um, I know neuroscientists have been trying to study, uh, another aspect of consciousness, which is the sort of a theory of the mind based one being that I think I'm summarizing this correctly. 01:08:19.42So it has to do with the fact that I'm able to anticipate what you're thinking and to anticipate your behavior in a situation I'm, I'm witnessing from, from the outside. 01:08:29.28And that's, there's this really very, very interesting experience with, with kids, um, at least, uh, and I hope to do that with, with mine because I find it very funny, which is that you're able to basically until they are three or four years old, they are not able to understand that someone may, may not be seeing, may not be aware of what they are seeing. 01:08:48.78So if, if, if someone enters a room and, and, and, uh, if we know that if someone wasn't inside this room where I, when I hid an object, we know that that person wants, will not be able to find the object. 01:09:03.32A kid under three or four years old will not know that because he's not, he or she's not able to put him, he or herself in, in the other person's shoes. 01:09:11.08Uh, and that's, that's, that's, that's a feature of consciousness. 01:09:14.28So that's the, it's one of the aspects of consciousness for me to be, to be able to know that I'm aware of something, you can be aware of it too, but our two awarenesses are, are different. 01:09:25.24Uh, that's one aspect of consciousness, uh, which you can study directly, I think. 01:09:30.08And like the other ones, that's what I mean. 01:09:31.76So this is, because it changes with, with, with one getting older. 01:09:35.86So you can study a kid, uh, as he evolves from state A to state B of consciousness in this, in
Gautam Agarwal 01:09:43.42this regard. 01:09:44.42And, uh, what would you say we could learn from that? 01:09:48.48And let's say, so that the child gets this theory of mind, like, what could you say about consciousness from that or, or anything else? 01:09:55.12No, you could say maybe how it, how it emerges from, from the brain.
José Maria Pimentel 01:09:59.52So by studying the way the brain changes from, let's say three to four years old or four to five years old, you'd be able to ideally pinpoint the, the, the origin of that change in, in consciousness. 01:10:13.26Yeah.
Gautam Agarwal 01:10:13.74This is probably much more complex than that. 01:10:15.64Well, um, so, um, it makes me wonder, for example, is it a, uh, a change at the level of physiology, connectivity, or is it a conceptual change, a set of a change in the rules that through which you view the world, which might be a type of change that is much harder to detect? 01:10:36.62Okay.
José Maria Pimentel 01:10:36.84I would say my, my bet would be that it, it happens at the level of neurophysiology. 01:10:42.30So at the level of the actual brain, because if it was merely conceptual, it will not happen at the same age with everyone.
Gautam Agarwal 01:10:49.74I don't know if it happens at the same age with everyone though. 01:10:53.02Because, uh, at least my limited experience with my one year old is that there's, um, a little bit of fuzziness in terms of, uh, the exact time. 01:11:03.58What, what seems to be, I don't know, like, uh, and once again, this is kind of anecdotal with just saying him, there's kind of a chain of, uh, of events. 01:11:12.48Like first, um, I, I learn how to, uh, crawl and then I learn how to prop myself up. 01:11:20.18There's these kinds of antecedent behaviors that open the door for the, for the next behavior. 01:11:26.54And, uh, for, so, you know, I guess what you're saying raises the question with any behavior that a child goes through, is it, is it the same sort of thing in terms of your point about this theory of mind? 01:11:40.40Uh, to me, one issue with that is that it, um, raises this, um, distinction between the so-called easy and hard problems of consciousness. 01:11:50.68Are you, have you heard of that? 01:11:52.00So it's, uh, this guy, David Chalmers, uh, came up with this notion to try to distinguish, um, ways in which we think about consciousness. 01:12:03.00And he, what he's saying is that, um, many of the phenomena that we ascribe to consciousness actually fall into the so-called easy problem, which is, uh, more functional in the sense that, um, it's, uh, you can train an AI to say that, okay, now, there, uh, you can have an AI that has a label, a flag that it raises every time it is, uh, confronted with another agent that is in the room and having the same experience. 01:12:34.16So you can train a system that, uh, responds to sensory input to produce a certain behavior. 01:12:41.34And many of the things that we do and that we're capable of actually fall into this kind of easy problem of consciousness, which is, uh, ultimately a cognitive issue. 01:12:52.48And what he thinks of as the hard problem is the phenomenon of, uh, a certain sensory experience. 01:12:59.48And so, um, curious about your views. 01:13:01.98Do you see this theory of the mind? 01:13:04.28To me, it seems like it could be seen as more of the, the kind of easy rather than the hard problem, but I don't know if, if you agree with that.
José Maria Pimentel 01:13:10.82Yeah. 01:13:11.12I mean, I'm glad you asked my opinion because I, I'm the layman here, but I, but I think it's, I would put it in the first one too. 01:13:18.04And actually I've heard about that before. 01:13:19.82Um, I just said, heard it by, by different names. 01:13:23.64So the, the, the hard one was called phenomenological and the first one had a different name. 01:13:28.48I don't recall. 01:13:29.22Um, but yes, and actually I think when we, most of the time when, when we talk about consciousness in a more philosophical way, we, we, we're talking about the second one, about the phenomenological one that has to do with like human experience and, and sort of, um, you reflecting upon your own experience rather than the first one, which is more of a more, I don't know, more operational one. 01:13:53.52And if, if you will, this theory of the mind might be between the, I think it's kind of a bridge between the two of them. 01:13:59.68So that's why it makes sense for it to occur, uh, when we are growing up.
Gautam Agarwal 01:14:03.18So during our development, do you think babies are conscious and in the, in the hard case? 01:14:08.36I was going to ask you about that. 01:14:09.74Okay.
José Maria Pimentel 01:14:10.48I don't know. 01:14:11.22Actually, I don't know. 01:14:12.30I don't know where is baby, where, whether babies are conscious. 01:14:15.00I don't know. 01:14:16.06I have no idea when we started to be conscious. 01:14:18.70Uh, so we human beings. 01:14:20.12So there, there's a theory that it happens. 01:14:22.06It happened actually not too long ago. 01:14:24.12So it happened like with the agricultural revolution. 01:14:26.96Uh, and that would tell us that it's more of a software based thing there and a hardware based thing. 01:14:32.84Uh, for instance, are, are dogs conscious? 01:14:35.64I don't know. 01:14:36.22What do you think?
Gautam Agarwal 01:14:37.08Um, I would place money on yes, but it depends on what we define, right?
José Maria Pimentel 01:14:42.46Uh, I think they are, they clearly are, they clearly have emotions. 01:14:47.06So that's the, they have that clearly. 01:14:49.30They do not, they do not, they do not have this higher level of consciousness of being able to, to think about themselves and to reflect the, because they are not, their brains are not, are not, are not developed as that, but they, uh, and if you go from the dogs are not even the most intelligent, uh, beings, um, below human beings. 01:15:08.84But if you, if you start going, if you start from human beings, uh, no monkeys, dolphins, going out of the dogs, and then, and then going all the way, uh, down, where's the border between soft consciousness and no soft consciousness. 01:15:21.92There are a lot of, um, uh, ethical challenges there because if, if one decides to be, to be vegetarian, because we don't want to, to hurt conscious beings, uh, implicit in that is our judgment of what's, which beings are conscious and which beings are not. 01:15:39.08So it's, that's, that's, that's all very tricky. 01:15:41.52Uh, I find this very fascinating, this whole topic of, of consciousness. 01:15:46.48So I'll close this one. 01:15:48.00So I think we've, we've talked enough about consciousness. 01:15:50.64We almost figured it out. 01:15:51.90Well, yes, we did nothing like that. 01:15:55.40That's what we expected. 01:15:57.34Let me go, let me, let me go back to the, let me try and close the other two topics we, we were jiggling with. 01:16:05.00So, um, if, if I can do it, if I can classify them this way. 01:16:08.76So one, one more about the brain and the other more about AI and, and, and the comparison with our intelligence and our, the way the human brain thinks and, or operates and the way AI has done so far. 01:16:19.88So on, on the first one, there are some, some doubts I've always had and I cannot, I mean, I, I, I would be stupid to let, to let it go away without asking a few questions about the brain I found very interesting. 01:16:31.08So one, one of them, I actually, I heard one neuroscientist, I think talking about this once and it had never occurred to me, which is why the hell do we have two hemispheres in our brain? 01:16:41.52I mean, why, why having a bridge where you could have the cities, the two cities merge?
Gautam Agarwal 01:16:48.80Yeah. 01:16:49.52So, um, well, so like the, in English you call it the cheeky answer. 01:16:55.04I don't know what, what you would call it in Portuguese. 01:16:57.02I have no idea. 01:16:57.60The kind of, um, very simplified answer is, uh, why do we have two hands? 01:17:02.98Um, but I would say a more serious answer would be, um, in a way related to the idea, the utility of off. 01:17:10.62And actually there's, um, a neuroscientist, uh, Giulio Tononi, um, who has developed a, uh, one of the few quantitative, uh, models of consciousness. 01:17:23.70Um, so to try to give a prediction of a number associated with the degree of consciousness of
José Maria Pimentel 01:17:31.22a system, which, sorry, which, uh, consciousness brought sense or, or type A or type B?
Gautam Agarwal 01:17:37.48Oh, um, good question. 01:17:40.58Uh, I will not answer that for the moment because I, I don't know. 01:17:44.74Um, but it's not too important. 01:17:46.72All right. 01:17:47.22It, it could tie to both though. 01:17:50.16Uh, and what he says is that imagine that, um, a system is a bunch of nodes and the nodes are connected in a certain way. 01:18:02.38Um, if you don't, if you don't have many connections across these nodes, then the system is stupid. 01:18:08.70You don't have any potential to generate, uh, some sort of group intelligence. 01:18:16.30Um, on the other hand, if the system is completely connected, every node influences the others, uh, very strongly, then it's also stupid. 01:18:25.02It's the colon of a seizure. 01:18:26.84Everything is going on and off together. 01:18:28.68It has no, no criterion. 01:18:29.98Exactly. 01:18:30.62You need boundaries or you need, uh, disconnectedness in the system for it to be, um, useful. 01:18:37.44And, and likewise, one phenomenon that seems to exist in, in the brain, both at the level of hemispheres and the level of different regions is the specialization, the ability to actually, um, segregate the, um, activity in one region from another. 01:18:55.34So you might've heard, for example, language is more specialized in the left hemisphere, for example.
José Maria Pimentel 01:19:01.30Um, but is, is that true? 01:19:03.10So I was, I was, I was had that, that curiosity. 01:19:05.96Is that an oversimplification to talk about the left, for instance, left brain, uh, left side of the brain and the right side of the brain? 01:19:12.70Or does that, does that have, uh, uh, uh, actually a neuro neuro physiological basis? 01:19:19.04Yeah.
Gautam Agarwal 01:19:19.60Um, I think it's a, to me, it's a, it's a real puzzle, this idea of the global nature of the brain versus the local nature of the brain. 01:19:28.32And there's, um, evidence supporting both. 01:19:31.54So for example, you can have strokes in relatively specific regions, people who are undergoing surgery and have certain areas and brokas damaged in, in one hemisphere can have very, uh, severe effects. 01:19:45.84Um, so in that sense, the brain is very local. 01:19:49.36Um, at the same time, if you look at, um, the activity of the brain, let's say in response to a word or an experience. 01:19:58.12Happens everywhere. 01:19:58.90Exactly. 01:19:59.26Um, and, and so the, the closest kind of analogy that I've come to make of all of this is something along the lines of, um, uh, once again, I'll risk oversimplification, but cities in a country. 01:20:12.22So you take the U S for example, you have all of these cities that operate in roughly the same way, like Los Angeles, DC, New York, a bunch of people consuming a lot of energy and sending out a lot of products. 01:20:26.80But each one of these cities has a certain specialization, politics, finance, uh, art, and they gain something by this kind of segregation while actively participating in this, um, network. 01:20:40.10Um, so that's the way I'm thinking about it now. 01:20:42.48In fact, I think one of the, the current trends in, in neuroscience and, and, um, draws a lot from mathematics is ways of thinking, uh, quantitatively about highly connected large systems. 01:20:56.16We are for the first time, uh, in a position to record from thousands of neurons, uh, the entire brains of certain creatures to produce data when every single neuron is active or inactive to, to, to really get at this issue of, uh, what exactly is the utility of segregation? 01:21:15.44What is it to be functional? 01:21:16.58In a way, just because I damage something and it prevents you from speaking doesn't mean that it's causing your speech. 01:21:23.66It just means that it's, uh, a critical point in this whole network.
José Maria Pimentel 01:21:28.22We've done that with worms, right? 01:21:30.24Yeah. 01:21:30.56Yeah.
Gautam Agarwal 01:21:31.20The story that I've heard at least is that the worm is kind of a so-called solved system, at least at the level of connectivity. 01:21:38.74Yes, exactly. 01:21:39.46Um, now creatures like the worm are very notorious because, uh, once again, for evolutionary reasons, they, um, they're heavily multiplexed, meaning the same connection can release multiple chemicals at different timescales. 01:21:55.10Um, and, and so just because you know where a connection is, especially when the creature is really small, as far as we can tell, uh, it doesn't actually give you much insight in terms of the functional nature of that interaction.
José Maria Pimentel 01:22:09.68Yeah, that's very interesting. 01:22:10.60In our case as, as human beings, that, that there's another aspect, which I'm, I haven't heard it mentioned, I think, anywhere when talking about AI, but, but one specific specificity of our body and of our nervous system is that it's, it is not, it is not circumcised, it's not only in our, in our head. 01:22:29.62So we have our nervous system extends to the, to the rest of our body. 01:22:33.04So we have neurons in the rest of our, of our body. 01:22:36.06I mean, despite this, is it fair to say that we only think inside the brain or can we, can we actually extend the thinking process to the, towards the, the, the whole body?
Gautam Agarwal 01:22:45.36Yeah. 01:22:46.52And can we extend the thinking process across bodies? 01:22:51.06Like for example, in this conversation, uh, we are to some degree controlling each other's thoughts and influencing the direction of where you're going to go. 01:23:02.46Um, in terms of what you raised, I think, uh, the gut is a good example. 01:23:07.88Like if you look at the, exactly, it's heavily innervated. 01:23:12.54And I think many of us have had the experience when we're nervous that we feel something that seems to come from, from this region. 01:23:20.38Um, another thing that I imagine with, with training, one can get more, more access to as well. 01:23:27.72Um, but in, in some ways I feel like your question goes back to the, this philosophical depth of, you know, what is thought and, you know, what is the substrate of thought?
José Maria Pimentel 01:23:38.60Yeah. 01:23:39.06I imagine that. 01:23:39.84No, but I was thinking in this case, let's say, so we talk about the case of having a stroke or being of, of, or having your brain damaged. 01:23:47.92And sometimes that, that, that can change the way you behave. 01:23:51.00Could we say the same thing about the gut, for instance? 01:23:53.32So having the, having some damage to your gut, would that feed into, into a different behavior? 01:23:58.46Or is that only the receptor? 01:24:00.64Is that only at the receiving end, end of your cognition or does that feed into your cognition too?
Gautam Agarwal 01:24:06.36Yeah. 01:24:06.96Well, I've heard ulcers are very painful and distracting. 01:24:10.24Um, so that would be kind of a gross example, but then, uh, more recently, there's also a lot of discussion about the microbiome, um, the, uh, community of bacteria and, and other microorganisms that live in your gut that apparently can influence your mood. 01:24:30.36A lot of microorganisms interact with some of the same neurotransmitters that, that we do. 01:24:36.12They metabolize many of the same molecules we do. 01:24:38.74So to me, it's no surprise. 01:24:40.44And I'm really looking forward to where that goes. 01:24:44.42It could be a very deep and profound level of effect. 01:24:47.76Like, does it have influenced the, the kinds of thoughts that I have? 01:24:50.70Like happy thoughts, sad thoughts, deep thoughts, uh, it could be, uh, influenced by this.
José Maria Pimentel 01:24:56.14No, and actually this is a good segue to go to, to go back to AI, because one of the things I've, I've heard as being one of the limitations of AI, and this, this sounds, this sounds like a very trivial thing, but it's the fact that the, an, an, an artificial intelligence program doesn't have a body, meaning that there's no, there's no way to, to interact with the environment. 01:25:16.26We, and we are to a large extent the way we, we learn by interacting with the environment. 01:25:20.70So I, we, we, we learn by, by, by, well, we have five senses obviously, and we, we interact with each other and we need, we need a body to collect data points of those experiences, which, which, uh, then we'll feed into our learning process. 01:25:35.54Uh, and if, I mean, if, if, if there's no, if, if, if, if, if we don't find a way to feed an AI system that's analogous to these, um, maybe that's a huge limitation, right?
Gautam Agarwal 01:25:47.74Yeah. 01:25:47.98And in fact, many, um, practitioners of AI, they spend a lot of time, uh, thinking about what is the right way of representing the input and the output. 01:26:00.88Um, so the same architecture can do very well or very poorly depending on how you massage the inputs and, and the outputs. 01:26:11.08And, and as you're alluding to, there's a lot of intelligence at the level of our sensory motor, uh, effectors. 01:26:18.36So for example, the eye, it has all of this adaptive optics that is very, um, precise. 01:26:24.86And, and many of the senses, they respond at these physical thresholds. 01:26:31.88So for example, the ear can detect, um, sounds that are basically at the level of atomic vibrations, um, or at, at room temperature. 01:26:40.96So just fluctuations, thermal fluctuations. 01:26:43.76Our ear? 01:26:44.54Yes. 01:26:44.98Really? 01:26:45.26Yes. 01:26:45.82Uh, and so in that sense, uh, the, uh, at the level of the musculature, the muscles have all of these kind of elastic properties that, uh, allow, for example, a cat with no connection to its brain walk on a treadmill. 01:27:02.30There's a certain springiness to our bodies that, uh, allows itself to stabilize in a certain way that might make the computational problem that the brain is solving much simpler in, in a sense. 01:27:15.04Um, now we've learned a lot from physiological studies and you can try to input that into, uh, the so-called reinforcement learning environment. 01:27:24.86I was describing earlier as a set of filters that kind of, uh, transforms your data. 01:27:30.64But one, one thing that I, I believe the embodied AI research community is finding is that, um, they learn when they try to put an AI in a body, they learn all of these unanticipated aspects of physics that, that are very fundamental to the problem that weren't seen before. 01:27:51.22Yes. 01:27:51.66And you only discover them when it's too late. 01:27:54.12But have they done that? 01:27:55.22Well, in the sense of, uh, robots that, that have like, uh, you know, towel folding robots, things like that, um, if you try to make a navigating robot, often you'll find, um, various aspects of how it collides with objects or certain types of surfaces like carpets, for example, that we might take for granted may represent a certain challenge that you don't think of until you, you implement it.
José Maria Pimentel 01:28:22.00Yeah. 01:28:22.06That's, that's quite interesting. 01:28:23.64We've covered some of the limitations that are usually mentioned in, in terms of AI converging with, with human intelligence. 01:28:32.44And this is very funny because when we say converging, uh, we're not talking actually about converging, but if that ever occurs, it will obviously surpass human intelligence because if, if everything is in place, as we mentioned before, the, the processing power of a computer is so much larger than our own, that it will obviously surpass. 01:28:50.16So that's, this is kind of a binary thing. 01:28:52.24There will never be equivalence between human and AI. 01:28:55.42Either AI is below or, or AI is above here, which is, which is a, which is a funny insight from, funny, I mean, curious insight from, from all this. 01:29:04.50Um, but another aspect that's a very colloquial one that's often pointed out is the, is the fact that AI lacks common sense, what you can call common sense, which is, which is funny because we talked before that AI in the, in this very complex, in one sense of the word games, uh, completely beats the hell out of us. 01:29:24.72So in, in, in, in, in, in chess, in, especially if you go to, to, to more simple games, we, we lose every, in every round, but then, but then there's some degree of what we, we call common sense. 01:29:37.56And the word common tells us that everybody has that regardless of the IQ of the, of the individual, they have it or most likely they have it. 01:29:45.04And AI then likes that. 01:29:46.50So what's that if not complexity? 01:29:48.34So it's, it's, it's, again, the word complexity is very tricky to use, to use in this regard, but it, one case of the, one case could be the, the traffic sign example, uh, or I don't know that there's another example, which is, I don't know if that can be called common sense, but when you, let's say I used a word with a letter, which is halfway between an A and an H. 01:30:11.34If I use that letter in, in, in two words, in the first word, it's, it's representing an A and in the second word, it's representing an H. 01:30:17.36You will get what I'm writing and I will obviously get it. 01:30:23.02No problem for AI. 01:30:24.22That's very tricky because they are the same word. 01:30:26.80Uh, I'm not sure that's, that qualifies as, as, as common sense.
Gautam Agarwal 01:30:30.62Sure, sure. 01:30:31.02I mean, what you're referring to is kind of a sensitivity to the context in which an experience happens. 01:30:36.98Uh, one aspect that, um, to me is intriguing is that we are judging, um, AIs in a human world and like, what would happen if we judged humans in an artificial world? 01:30:51.18Uh, would, would the opposite happen? 01:30:53.44Would we lack the so-called common sense of this digital world? 01:30:57.18Well then, in fact, um, a very cool study that has been done, um, more recently is, um, using, uh, looking at AIs and humans play video games. 01:31:08.90And, um, as we were saying before, it takes a while for, um, humans to learn or AI to learn how to beat these games. 01:31:17.66But what, um, this group did is that they took, um, a game and they replaced all of the familiar characters with a scrambled set of pixels. 01:31:28.02And now you get the human to play the same game or they took very basic assumptions that we make about physics, such as if I move to the left, I'm going to end up in the next pixel. 01:31:40.38Uh, they scrambled physics. 01:31:41.88So now if I move to the left, I could end up totally different place of the, of the screen. 01:31:46.02And so now, um, as far as the AI is concerned, all of these are the same because it's just getting patterns and making sense of those patterns. 01:31:54.38So it does equally well. 01:31:56.22And all of the privilege that we get from our common sense by playing games that other humans have designed who share our architecture, we lose all that instantly as soon as some of these, um, agreements are taken away from us.
José Maria Pimentel 01:32:11.22Yeah, because they came from our world. 01:32:12.58That's, that's one problem in learning. 01:32:14.58For instance, if one studies, um, these very fundamental branches of science, uh, such as physics, for instance, that there's, there's, there's part of the, of the way which you can do intuitively. 01:32:30.66So going back to intuition, uh, but then at some point you have to do it only by using mathematics. 01:32:36.32And that gets very, very unnatural for us because our brain is, is that's why it's so easy for me to, to explain things to you and vice versa by, by, um, calling your intuition and giving common examples or, or telling an anecdote or, or telling a story. 01:32:53.36But mathematics is, it's, it's very unnatural to us. 01:32:56.38So we, we can use it when we go to, from intuition to, to mathematics, but, but the way back is very tricky for us, for us human beings.
Gautam Agarwal 01:33:05.04Yeah. 01:33:05.92What I find curious is those cases of people who study things like mathematics, which as you're saying is very abstract, but they study it for decades. 01:33:14.60And it seems like those people, some of them, at least they develop a certain common sense about that. 01:33:21.24They can even visualize what is it to be in five dimensions or what does it mean for information to propagate in this kind of network? 01:33:30.16They have a kind of felt sense about these things.
José Maria Pimentel 01:33:32.76No, that's exactly. 01:33:33.74That's what I'm pointing at. 01:33:34.58I think, I think there's, even when we are abstracting mathematically, we always need to have a abstract that that's intuitive and those like very, very good and very, how can I put it? 01:33:52.36A very, very developed mathematician, mathematicians, they need to, they probably build their own intuitions as you were alluding to. 01:34:00.80And that might be, it might be interesting to study them. 01:34:03.04So to understand how they have built intuitions to, to think about realities that don't exist even in the real world, but they have, they, they, even those realities, they, they felt the need for them to be able to, to think about them and to process that information and to, to, to, to, to understand it. 01:34:20.98They, they, they needed to, to build their own intuition. 01:34:24.34So they were kind of imagined realities to understand that even in those extreme scenarios.
Gautam Agarwal 01:34:29.34And it seems like some of these, uh, properties might be due to the architecture of the brain itself. 01:34:35.74So, um, mathematician, there's a project which follows, uh, exceptional mathematical ability over, uh, the development of, of children and, uh, to adulthood. 01:34:48.14And you see that actually some of the science for, uh, mathematical genius can be seen in people at the age of 10 or so. 01:34:58.00So even at that age, there's something going on in these people that will facilitate their brains, uh, towards these, uh, mathematical, uh, exceptional mathematical abilities. 01:35:09.94And that suggests to me that it might be, um, somehow the architectures of their, of their brains that.
José Maria Pimentel 01:35:17.00I think I'll, I'll finish this. 01:35:18.84Yeah.
Gautam Agarwal 01:35:19.24There is a restless, uh, brain that is, um, eager to consume the world.
José Maria Pimentel 01:35:24.40One interesting example about, about that common sense and, and contexts of we were talking about, it was given by a very well-known philosopher, Ludwig Wittgenstein. 01:35:34.96I never know if I pronounce his name well, when he talked, he talked about a thing called, well, his concept was family resemblance, which I find very interesting. 01:35:42.28It can, it can be applied to a lot of things. 01:35:43.80His example was games. 01:35:45.26So we all know what, what a game is, but there's no single common feature across games. 01:35:51.56So you cannot say that the game is about scoring points because there are games that are not about scoring points. 01:35:57.52You cannot say that the game is about being faster than the other because there are not, some games are not about that. 01:36:03.00You cannot say that games are about playing in the team because sometimes are not, some games are not about that. 01:36:08.52But in the end, we all know what the game is because there are some family features as as he calls them. 01:36:13.42So that, I think that relates to the, to the common sense. 01:36:16.12But anyway, so we are wrapping up with our new guests from this last segment of the conversation. 01:36:21.58Let me just, we'll go to the book, but let me just ask about something I was about to forget. 01:36:26.80I cannot let you go as a neuroscientist without asking this. 01:36:30.30Why do we dream? 01:36:31.26It's a very simple question, but I think, I think it's, it's, it has always been a puzzle evolutionarily, but even for, I guess for, for someone who studies the brain, I think you may have some insights on this. 01:36:43.38I mean, why being asleep for one third of the day? 01:36:46.12So why could our brain work without, without us sleeping?
Gautam Agarwal 01:36:50.88That's a, that's another one that I would say at the level of what is consciousness. 01:36:56.14It's a very hard problem. 01:36:58.10It is one that I would say is maybe a little more accessible than the consciousness problem in a way. 01:37:04.48One idea is that dreams are, from the neuroscientific perspective, one idea is that dreams are ways of synthesizing information. 01:37:16.66So you get a lot of sensory evidence and it can be kind of piecemeal in nature.
José Maria Pimentel 01:37:21.56Kind of putting the house in order in a sense. 01:37:23.86Exactly.
Gautam Agarwal 01:37:24.00Yeah. 01:37:24.24And in fact, it's inspired AIs that actually operate on this principle as well. 01:37:30.06And it seems to really, to help their learning process.
José Maria Pimentel 01:37:32.28Oh, I didn't know that. 01:37:33.06Yeah.
Gautam Agarwal 01:37:33.46The so-called sleep-wake learning cycle. 01:37:37.34I personally am not convinced that that's the entire story based on my own experience, just because some of the dreams that I and others have had seem to give me really surprising insights into my personal life. 01:37:56.28and seem to be, I've even had a dream that seemed to predict something that actually happened in my life in the future. 01:38:04.16And I can't really make sense of that. 01:38:05.96Yeah. 01:38:06.06So I'll close, I guess, by saying that I look forward to finding out.
José Maria Pimentel 01:38:12.16You close with a, with a, with a, with a, with a, almost a teaser. 01:38:16.84Exactly. 01:38:17.48Because that seems like a very interesting story. 01:38:19.48Uh, and you can interpret, I mean, obviously can interpret it in a transcendental way, but you can interpret it in terms of your brain finding patterns that you're, that you're not consciously aware of. 01:38:28.78But anyway, I've had you in this, in this inquiry long enough.
Gautam Agarwal 01:38:32.44So let's go for the book. 01:38:33.46The book that I chose, um, is, um, called Psychotherapy East and West by Alan Watts. 01:38:40.96And it was published in the, the early seventies. 01:38:44.30Uh, Alan Watts is someone who's quoted, um, often in, um, you know, various songs and YouTube videos. 01:38:50.98Um, it turns out that he's, um, a scholar who's read very deeply both in, um, psychoanalysis over the 20th century, as well as Zen Buddhism. 01:39:01.42And so he draws from these two, um, lineages of, of inquiry into the workings of the mind. 01:39:10.24And in this book, he tries to get at the question of how do Eastern and Western approaches of psychology help us in, in our, in the problems of man. 01:39:23.56And, and what he posits is that, um, both of these traditions, the Eastern and the Western tradition, um, human, the human being as someone who is trapped in a game with no answer. 01:39:40.46Um, so because we are living in a world of contradictions where we are citizens that are free to behave how we like, and we are expected to have responsibility. 01:39:52.78This would be an example of a contradiction in some sense. 01:39:56.44We are doomed to not find a unifying answer. 01:40:00.62And so in his mind, both of these processes working with a therapist or in a Zen practice, allow you to see the contradictions of the game that we're embedded in and recognize it as just that, thereby freeing us from, from the contradictory chains of existence.
José Maria Pimentel 01:40:20.24I'm looking forward to reading the book. 01:40:21.94I was actually checking it. 01:40:22.80It's from 1975. 01:40:24.38So it's, it's, it's a, it's a classic almost. 01:40:26.80Great.
Gautam Agarwal 01:40:27.40Okay, got them. 01:40:28.20Thanks a lot, man. 01:40:29.02Thank you so much.
José Maria Pimentel 01:40:36.16Gostaram deste episódio? 01:40:37.70Se encontram valor no 45 graus, existem várias formas de contribuir para a continuidade deste projeto. 01:40:42.90Podem avaliá-lo na aplicação que utilizam, seja ela o iTunes, Spotify, Stitcher ou outra, e podem também partilhá-lo com amigos e comentá-lo nas vossas páginas ou redes sociais. 01:40:51.46Se acharem mesmo que merece e puderem fazê-lo, podem ainda tornar-se mecenas deste podcast através do Patreon ou do PayPal. 01:40:57.76Com esse apoio, estão a contribuir para a viabilidade deste projeto, que passa a ser também um bocadinho vosso. 01:41:02.56Para além disso, obtêm em troca vários benefícios, como, por exemplo, o acesso ao backstage do podcast e também a possibilidade de sugerir perguntas aos convidados. 01:41:11.52No fim do dia, já se sabe, são os ouvintes que tornam possível um projeto destes. 01:41:15.06Ouvintes como Gustavo Pimenta, João Vítor Baltazar, Salvador Cunha, Ana Mateus, Nelson Teodoro, Paulo Ferreira, Duarte Doria, Gonçalo Martins, entre outros mecenas, a quem agradeço e cujos nomes podem encontrar na descrição deste episódio. 01:41:28.28Até à próxima. 01:41:32.16Legenda por Sônia Ruberti