In this Directors’ Conversation, HAI Denning Co-Director Fei-Fei Li’s guest is William Newsome, the Harman Family Provostial Professor of Neurobiology at the Stanford University School of Medicine and the Vincent V.C. Woo Director of the Wu Tsai Neurosciences Institute.
Here Li and Newsome discuss the role of computation in neuroscience, the challenges computational neuroscientists can address, whether understanding the brain at a molecular level can lead to better neural networks, AI’s motivation spectrum, and the complicated definition of consciousness when it comes to both natural intelligence and artificial.
Fei-Fei Li: Welcome to HAI’s Directors’ Conversations, where we discuss advances in AI with leaders in the field and around the world. Today with me is Professor Bill Newsome. I’m very, very excited to have this conversation with Bill. We have been having conversations throughout our career, and I’ve been such a great admirer of Bill’s scholarship and leadership. He’s the Professor of Neurobiology at Stanford School of Medicine and the Director of the Wu Tsai Neurosciences Institute at Stanford. Bill has made significant contributions to understanding of neural mechanisms underlying visual perception and simple forms of decision-making. As head of Wu Tsai Institute, he’s focused on multidisciplinary research that helps us understand the brain, and provide new treatments for brain disorders, and promote brain health.
So welcome, Bill. I’m very much looking forward to this conversation.
Bill Newsome: Great to be here, Fei-Fei. Nice to see you again.
Li: Let’s start with just defining and talking about the intersection of AI and neuroscience. What do you see as the role of computation in your field, and Wu Tsai Institute’s work?
Newsome: Well, that’s a great question. Computation is extremely important in the field of neuroscience today. There are two or three different ways I could answer that question, but let me try this one on you. We actually have a subfield called computational neuroscience. We’ve hired faculty in this area here at Stanford, and we hope to hire more. People sometimes ask me, “What is that?” And I would put it this way. Computation in neuroscience has about three different areas that are really, really important. The first area is theorists: it’s people who actually try to theorize and extract general principles about how the brain is computing, how the brain is representing, how the brain produces action.
The second area I would call neural network kinds of people: modelers, people who really understand how to do deep convolutional networks and understand how to do recurrent neural networks, and actually model simple toy problems that we know the nervous system solves. If we can figure out how these networks solve these problems, then we can get some insight maybe into new hypotheses about the nervous system.
And then a third way that computation is really influencing neuroscience is through high-end data analytics. Like many areas of science, we are getting larger, richer, more sophisticated, and sometimes much more obscure datasets now out of the brain than we’ve ever been able to get in human history. And to actually understand how to deal with those data, how to treat them, how to avoid statistical pitfalls is extremely important.
I think all three of those areas of computation are really important for neuroscience. I think different computational neuroscientists may excel in two, or occasionally even all three of those, but we need more people like this in neuroscience today, not fewer, because the challenges are greater than they’ve ever been.
Li: What are the biggest challenges that you feel are in need of this kind of computational neuroscientist?
Newsome: Well, let me give you some real-world examples; maybe give you one where computation has actually played a leading role and other approaches to neuroscience are lagging behind, and one where computation needs to step in and create a role in order to create understanding.
So one area that I would give to you is something in the nervous system called integration, and it’s familiar to anyone who’s taken calculus. It’s literally counting up events that happen, literally integrating some time series and saying how much you have at the end. This turns out to be a really important problem in the nervous system in many areas, including decision making, but a very simple one. It’s just moving your eyes. So we know when you move your eyes from this point to this point, certain neurons give a little burst in a map of eye movement space in the brain, and when they get the burst, the eyes go out there. The amazing thing is they stay there after they get there, even though the burst is gone.
And the theorizing, the theory of computation is about integration: How could you take neural signals of the burst, integrate some value that holds the eyes in position until an animal is ready to move the eyes again? We knew some things about the physiology, we knew some things about the computation, we knew some things about the connectivity of brain structures that produce eye movements. What we really lacked was actually anatomy, in this case.
It turned out that several different computational theories that embodied physical principles could account for this, but to know which one is actually working in the brain, we needed the microanatomy of how cells are really hooked up to each other. That’s an example where computational theory actually led the way and motivated some anatomical questions.
But one that many listeners to this will resonate with today is the example of deep convolutional networks that are approaching human performance in visual categor-
Li: Surpassing in some cases.
Newsome: Surpassing in some cases, absolutely. People’s jobs are insecure here because these networks are getting so good at visual categorization. And this presents a really interesting problem, because you’re going to train a deep convolutional network that can do these things that seem almost magical, and we know almost everything there is to know about them, right? We know exactly the connections between the layers of the trained network, we know the signals that are passing, we can see the dynamics that exist, we can measure their performance. But there is still this deep angst, an intellectual angst, I’d say, among neuroscientists, and I think among some people in the AI community, that we still don’t understand how that’s happened.
What are the algorithmic principles by which you take an array of pixels and you turn it into faces, and distinguish among faces? And somehow, the deep physical and computational principles are not there yet. We don’t understand how these things are working. We understand the learning algorithm, and maybe that’s as deep as it’ll get at some point. But here’s a situation where I think computation needs to step in and teach us this, both for the artificial networks, and then for the real networks in our brains that recognize faces.
Li: Bill, I want to elaborate on that because neural networks, especially in visual recognition, are dear and near to my heart. On the one hand it’s phenomenal, right? We have these hundreds-layered, sometimes even thousand-layered convolutional neural network or recurrent network algorithms that are just very complex and can perform phenomenally well. When it comes to object recognition, some of these networks do surpass human capability. But in the meantime, if you look inside under the hood of these algorithms, while they’re humongous, they’re also extremely contrived compared to the brain.
I’ll just take one example of the neuron-to-neuron communication. The way that it’s realized in today’s neural network algorithm is a single scalar value, whereas the synaptic communication in the brain, as we learn more, and your colleagues will tell us, is far more complex. The neural signaling is not just one kind of neural signaling. I would love to hear more about this.
Also, on a little more system level, our brain is this organic organ that has evolved for at least the past 500 million years; the mammalian brain is about 100 million years, and it has different parts, and different modules, and all that. And today’s neural network is nowhere near that kind of complexity and architecture.
So on one hand these humongous deep learning models are doing phenomenally well. On the other hand, they’re also very contrived compared to the brain. And I’m just very intrigued: from your perspective, as we learn more about these computational realities of the brain at the molecular level, at the synaptic level and the system level, do you see that we’re going to have different insights how to build these neural networks?
Newsome: I hope so, Fei-Fei. I think this is one of the deepest intellectual questions that computationally-minded neuroscientists argue about, and that’s to what extent are AI, and what I call NI — natural intelligence — going to converge at some point and really be useful dialogue partners? And to what extent are they simply going to be ships passing in the night, or they’re going to be parallel universes? Because there are these dramatic differences, as you point out.
One individual neuron — and our brain contains about 100 billion of them — is incredibly complex: incredibly complex shapes and incredibly complex biophysics, and different types of neurons in our brain have different types of physics. They’re profoundly non-linear, and they are hooked together in these synapses and ways that form circuits, and understanding and mapping those circuits is a big fundamental problem in neuroscience.
But something that should give all of us great pause is that there are these substances that are released locally in the brain called neuromodulator substances, and they actually diffuse to thousands of synapses in the space around them in the brain, and they can completely change that circuitry. This is beautiful, beautiful work by Eve Marder, who spent her career studying this neuromodulation. You take one group of neurons that are hooked up in a particular way, spritz on this neuromodulator, and suddenly they’re a different circuit, literally.
Li: Yeah, that’s fascinating. We don’t have that computational mechanism at all in our deep learning architecture.
Newsome: And another feature of brain architecture, that you and I have talked about offline together, is that brain architecture is almost universally recurrent. So area A of the brain has a projection to area B. You can kind of imagine that as one layer in the deep convolutional network to another layer. But inevitably, B projects back to A. And you can’t understand the activity of either area without understanding both, and the non-linear actions, the dynamical interactions that occur to produce a state that involves multiple layers simultaneously.
Many of us think today that understanding those dynamical states that are distributed across networks are going to be the secret to understanding a lot of brain computation.
I know that recurrence is starting to be built into some of these DCNs now. I don’t know where exactly that field sits, but that certainly is one of the ways you get dynamics.
Dynamics are, again, another universal feature of brain operation. They reflect the dynamics in the world around them, and the input but also the dynamics in the output. You’ve got to have dynamical output in order to drive muscles to move arms from one place to the other, right? So the brain is much richer, in terms of dynamics.
Another thing about the brain is it operates on impressively low power.
Li: I know, I was going to say the 20-watt problem. That’s dimmer than any lightbulbs we have. We hear about these impressive neural networks like GPT-3 or a neural architecture search or burst, an image that algorithms are all burning GPUs much more massively.
So how do you think about that?
Newsome: Well, I don’t think about it very much, except that our contrived devices are very, very inefficient and very wasteful.
We have a colleague at Stanford, Kwabena Boahen, who studies neuromorphic engineering, and trying to build analog circuits that compute in a much more brain-like way. And his analog circuits certainly are much, much more efficient in power usage than digital computers. But they haven’t achieved nearly the level of impressive performance and the kinds of sort of cognitive-like tasks that DCNs have achieved so far. So there’s a gap here that needs to be crossed.
Li: Yeah, I think this is a very interesting area of research. You mentioned the word cognitive, and I want to elaborate on that because I know we started talking about computational neuroscience, but cognitive neuroscience is part of neuroscience, and also in the field of visual where I sit.
First of all, half of my PhD was cognitive neuroscience. Second of all, in the past 30 years, I would give cognitive neuroscience a lot of credit in the field of vision to show to the AI world what are the problems to work on, especially the phenomenal work coming from the 70s and 80s in psychophysics by people like Irv Biederman, Molly Potter, and then getting to neurophysiology and cognitive neurophysiology, like Nancy Kanwisher, Simon Thorpe, showing us the phenomenal problem of object recognition, which eventually led to the blossom of computer vision object recognition research in the late 90s and the first 10 years of the 21st century.
So I want to hear from you, do you still see a role of cognitive neuroscience in, I guess, two sides of this: one is in today’s AI, which I think I have an opinion, but also AI coming back to help?
Newsome: I am not nearly as well versed or trained in cognitive neuroscience as you were. That was your graduate training. I think in a very simple-minded way about cognitive neuroscience, that may make our colleagues, may make you shudder, Fei-Fei, I’m not sure. I was trained as a sensory neuroscientist, trained in the visual system, the fundamentals of Hubel and Wiesel, and the receptive-field properties in the retina. And then the first processing in the brain, and then the cortex.
I was sort of getting into the brain, back in the 1970s and 1980s, thinking about signals coming from the periphery. We all called ourselves sensory neuroscientists, but there was another whole group of neuroscientists who were coming the opposite direction. They were having animals make movements: a right eye movement, like we’ve already talked about, or arm movements, and they’re looking at the neurons that provide input to those movements, and then they’re tracing their inputs back into the brain. And this was a motor science kind of effort.
And the sensory side and the motor side has enjoyed listening to each other talk, but they didn’t really talk about it very much. But they had to meet eventually. And I think one part of my career was playing a part in hooking those two things up. And we did it by studying simple forms of decision making. So giving animals sensory stimuli — that was my comfort zone — asking animals to make a decision about what they were seeing, and then make an operant movement. And if they got it correct, they got a reward.
Well, how did the sensory signals that are the result of a decision get hooked up to steering the movement? And that there, you’re squarely in cognition land. Some people refer to that as the watershed between sensory systems in the brain and motor systems in the brain. How do you render decisions?
You can think about sensory representations in the brain as kind of being like evidence, providing evidence about what’s out there in the world. But then you can think about these cognitive structures in the brain that have to actually make a decision, render a decision, and instruct movements. You can’t move your eyes to the right and to the left at the same time. Not going to happen. Sometimes you simply have to make decisions.
That’s how I kind of got into the cognitive neuroscience. And I think it’s one of the most interesting fields in all of neuroscience right now. I am hoping that AI and computational theory ... well, I know that computational theory is making contributions because some of the integration problems, integration of evidence from noisy stimuli, those kinds of theories, those kinds of theoretical models have deeply informed my own work in decision making. So computation theory are certainly making contributions.
I sometimes wonder about the other way around: What is that we are learning from vision and neuroscience that could inform AI? And you and I have had conversations about that as well.
Li: Right. So I’ll give you an example of a group of us, Stanford neuroscience people like Dan Yamins, Nick Haber, they are the young generation of researchers who are actually taking developmental cognitive inspiration into the computational modeling of deep learning framework. They are building these learning agents that you can think of as learning babies as a metaphor, where the AI agent is trying to follow the rules of the cognitive development of early humans, in terms of curiosity, exploration and so on, and learn to build a model of the world and also improve its own dynamic model of how to interact with this world.
I think the arrow coming from cognitive developmental science actually is coming to AI to inspire new computational algorithm that transcends the more traditional, say, supervised deep learning models.
Newsome: One example where neuroscience has really led the way for artificial intelligence and for convolutional networks and artificial vision is the deep understanding of the early steps of vision in the mammalian brain, where set field structures filtering for spatial and temporal frequencies have particular locations in space; the multiscale nature of that; assembling those units in ways that extract oriented Gabor filters. That’s typical of the oriented filter, typical in the early stages of cortex processing in all mammals. And that now is baked into artificial visual.
That was the first thing. You don’t even bother to train a DCN on those steps. You just start with that front end, and that front end came honestly from neuroscience, from the classic work of Hubel and Wiesel, as you know. Coming through some principle psychophysics and statistical analysis input from people like David Field. I think if I had to point to one thing that neuroscience has given to AI it would be the front end of a lot of the vision.
Li: That’s a really big thing, so absolutely.
Newsome: Fei-Fei, let me just say that the other challenges there — and I think yeomans of the young generation who are working on visual would acknowledge, I think everyone acknowledges this, really — is that the artificial visual systems even though they can surpass human performance in some cases after they’re trained, the learning process is so different for humans from the artificial systems.
The artificial systems need tens of thousands of examples to get really, really good, and they have to be labeled examples, and they have to be labeled by human beings, or what is your gold standard. Whereas I have this little 5-year-old daughter at home, and by the time she was two or three, she had looked at a dozen examples of elephants, and she could recognize elephants anywhere. She could recognize line drawings, photographs, different angles, different sizes, different environments. And she can play Where’s Waldo on the common children’s magazine. And this is profoundly different.
So here’s an example where human cognitive neuroscience and the study of visual development in young humans and young animals, I think, presents a real challenge for artificial vision, artificial intelligence.
Li: Yeah, I actually wanted to emphasize on a point you just made because it truly, using your word, is profound because the way humans learn biologically, your NI, natural intelligence system learn is so different. I still remember 20 years ago, my first paper in AI was called “One-Shot Learning of Object Categories,” but until today, we do not have a truly effective framework to do one-shot learning the way that humans can do, or few-shot learning. And beyond just training example-based learning, there is unsupervised learning, there is the flexibility and the capability to generalize, and this is really quite a frontier of just the overall field of intelligence, whether it’s human intelligence, or artificial intelligence.
Newsome: Yeah, I think both AI and NI have to be appropriately humble right now about this. We’re almost equally ignorant about exactly how that happens.
Li: In a way, I almost think it has a social impact for those of us who are scientists. We need to share with the public about the limitations because the hype talk of AI today, of machine overlord and all that, is built upon some of the lack of knowledge of the limitations of the AI system, and also the phenomenal capability of human intelligence to stumble.
Bill, I want to switch topic a little bit because I think what you are doing at Wu Tsai goes beyond some of these more lower level modeling. One of the most important charter mission of Wu Tsai is neuro disorder and healthcare-related. Here, I’m going to say something that I hope that you can even disagree on: should we view AI and machine learning more like a tool for our researchers and doctors, clinicians to use this modern tool of data-driven methodology to help discover mechanism of diseases and treatments? Are there any examples of work at Wu Tsai like that? Just in general, how do you view AI through that lens of studying neuro disorders?
Newsome: Yeah, that’s a really good question. Is AI really more of a tool to enable us to get on with the business of doing serious biology, or do the actual processes and algorithms and architectural structure of AI lend understanding to their correspondence inside the brain?
And I think the answer is both. So let me just give you a little Bill’s-eye-view of neuro disease. There are some neurological diseases that have psychiatric comorbidities, where the biggest problem is simply that cells in the nervous system, somewhere in the nervous system, start dying, for reasons we don’t know yet. Parkinson’s disease is an example, where it’s a particular class of cells, the dopamine arginic cells that start dying, and we don’t know why. And Alzheimer’s disease, cells start dying all over the brain. There’s some areas that are particularly sensitive but by the time an Alzheimer’s patient shows up in the clinic complaining of symptoms, they’ve already lost probably billions of nerve cells; certainly hundreds of millions of nerve cells by the time they become symptomatic.
And those diseases, I think, are going to be solved ultimately at a molecular and cellular level. Something is going wrong in the life of cells, and whether that’s in the metabolic regime, whether it’s in the cleanup regime, keeping the cell whole and safe and free of pollutants, whatever it is, the secrets to that are going to be in cell biology, and AI can certainly help us tremendously just by providing tools to assemble all the data that we’re acquiring at that genetic and molecular level inside of cells.
On the other hand, there are neural diseases that smack of the more systems type of pathology; the problems are not lying in single cells probably. So you take some of the symptoms of Parkinson’s disease, for example, the tremor and things like this, they can actually be rectified by putting stimulating electrodes inside the brain and doing a process called deep brain stimulation. Any of the listeners who aren’t familiar with this can just Google “deep brain stimulation,” or go to YouTube, and you can see amazing videos of remission of symptoms with this. Not a cure for Parkinson’s but it’s a treatment for the symptoms.
And there are things like depression, which themselves don’t kill people; it’s not like it’s a progressive degenerative disease. In depression, people come in, they come out. It’s a dynamic kind of process. It smacks of the state system inside the brain, and that state can go through multiple systems, some of which are depressed, some of which we would characterize as more normal or positive kind of outlook.
And that kind of dynamics of complex systems, I think, is going to be part and parcel of the AI computational neuroscience thrust: understanding how these densely interconnected networks, based on certain inputs, can assume different states, fluctuate between them. I think could give some insight into the actual disease itself.
So I think it depends on which disease you’re talking about, on whether AI is primarily going to be a tool or whether it might actually suggest some intellectual insights into the sources and explanations for some of them.
Li: That speaks of the broadness of machine learning AI’s utility in this big area. Where we sit at HAI, we see already a lot of budding collaborations between the school of Medicine, Wu Tsai Institute, and HAI researchers, where all of these topics are touched. I know there’s reinforcement learning algorithms in neurostimulation for trauma patients. Or there is computer vision algorithms to help neuro-recovery, in terms of physical rehabilitation. And also all the way down to the drug discovery, or those areas. So very excited to see this is also a budding area of collaboration between AI and neuroscience.
Newsome: I think that this is going to grow, that interface is going to grow. I think ultimately we’ll diagnose depression much better through rapid real-time analysis of language that people use, and adjectives that they use, than expensive interaction with physicians. I don’t think the algorithms are going to replace physicians, but they’ll be very useful.
Can I bug you about something I’m wondering about?
Newsome: In some of the first discussions that led up to the formation of AI that I was privileged to sit in upon, we raised a question of when you’re a biologist you think about a human or animal performing a task; doing a discrimination task, or making a choice between this action that action. And the question that comes up is motivation. What is the organism motivated to do at the time?
And this gets very complicated. In all kinds of social situations with humans, we worry about what’s fair, and we may do things that are against our economic interest because we are striking out for fairness. There are these values, there are these motivations, there are these incentives. And I wonder: what is the motivation in an artificial agent? To the extent that I know anything about motivation in artificial agents, it’s minimizing some cost function. Is that all there is to understand about incentives and motivation?
Is that all these complex feelings that we have, are they just reduced to cost functions, or is there a whole world there that AI needs to discover that they haven’t even scratched the surface of yet?
Li: This is a beautiful question. When you talk about motivation, I was thinking: what kind of reward objective mathematical functions I can write? And I come up very simple ones, like in the game of Go, I maximize the area that my own color block occupy. Or in the self-driving car, I will have a bunch of quantifiable objectives, that is: stay in the lane, don’t hit an obstacle, go with the speed, and so on. So yes, the short answer is motivation is a loaded word for humans, but when it comes to today’s AI algorithm, they are reduced to mathematical reward functions, sometimes as simple as a number, or what we call scalar functions. Or a little more complex, a bunch of numbers, and so on. And that’s the extent.
This clearly creates an issue with communication with the public because on one hand, people are claiming incredible performances of vision language, especially those confusing language applications where you feel the agent actually is talking to you, but the under the hood it’s just an agent optimizing for similar pattern it has seen.
So we don’t have a deep answer to this at all. My question to flip this is: both as a neuroscientist, as well as more objective observer of AI, do you see this as a fundamentally insurmountable gap, hiatus, between artificial intelligence and natural intelligence that would potentially touch on philosophical issues like awareness, consciousness? Or do you think this is a continuum of computation? At some point when computation is more and more sophisticated, things like motivation, awareness, or even consciousness would emerge?
Newsome: Well let me answer that in a couple of ways. First, I don’t think that it’s a fundamental divide. I don’t think that there’s anything magical inside our brains associated with the molecules carbon and oxygen, hydrogen and nitrogen. I do this thought experiment sometimes with groups where I say, “I’ve got a 100 billion neurons in my brain, but imagine I could pull one of them out and replace it with a little silicone, or name your substance, neuron that mimicked all the actions of that natural neuron perfectly. It received its inputs, it gave its outputs to the downstream neurons, can even modulate those connections with some neuromodulator substance, it can sense some: would those still be Bill with this one artificial neuron inside my brain along with a 100 billion natural ones?” And I think the answer would be yes. I don’t think there’d be anything fundamentally different about my consciousness and about my feeling. And then you just say, “Well what if it’s two?”
Newsome: What if it’s three? And you get up to 100 billion, after a while, and my deep feeling is that if those functional interactions are well mimicked by some artificial substance, that we will have a conscious entity there. I think it may well be that entity needs to be hooked up to the outside world through a body because so much of our learning and our feeling comes through experience. So I think robotics is a big part of the answer here. I don’t like the idea of disembodied conscious brains inside a silicon computer somewhere. I’m deeply skeptical of that.
Li: It’s like the movie, Her.
Newsome: Yeah. I don’t think the divide’s fundamental, I don’t think it’s magic. But the only place I know that consciousness exists, and that these intense feelings take place, is in the brains of humans, certainly in a lot of other mammals, maybe all other animals, but probably in birds and others as well.
But I think that neuroscience is kind of suited now, and artificial intelligence as it’s constituted now, there may be a fundamental divide just because they start with different presumptions, different goals of the kind that we’ve discussed here for the last half hour.
Does that make any sense to you, or am I just babbling here?
Li: No, no, it’s making some sense to me but let me try to share my point of agreement and disagreement. My point of agreement is that, like you said, where we are, the deep learning algorithm and also our understanding of the brain, is still so rudimentary. And from the AI point of view, it’s just so far, the gap between what today’s AI or foreseeable AI can do to what this natural intelligence from computation to emotion to consciousness, it’s just so far I really don’t see that the current architecture and mathematical guiding principle can get us there. What I don’t have an answer is when you say 100% of your neurons are replaced.
Newsome: But perfectly mimicking the functional relationships of the originals.
Li: First of all, I don’t know what perfectly mimicking means in that because we’re in counterfactual scenario. Like, maybe we can perfectly mimic up to this point of your life where your neurons are replaced, but what about all the future? Is that really Bill? It’s almost a philosophical question, that I don’t know how to answer. But I think this consciousness question is at the core of some of neuroscience researchers’ pursuit, as well as a very intriguing question for AI as a field.
Newsome: So consciousness, I call it the C word, and mostly I don’t utter the C word. But it is maybe the single most real, as Descartes thought, and interesting feature of our internal mental lives, so it’s certainly worth thinking about, both from a neuroscience point of view and an artificial intelligence point of view.
A lot of the muddiness about that word comes because we use it to mean so many different things. We use it to mean a pathological state, somebody’s unconscious rather than conscious. We use it to mean a natural state called sleep, and people are asleep and not conscious. Or we use it to mean: I’m conscious of this TV screen in front of me and I’m not conscious of the shoes on my feet at this particular point in time. Or we can use it at a much higher level, that I am conscious of the fact that I’m going to exist, that I’m going to die, that I have a limited time on this planet and I need to find as much meaning from those years as possible.
And so you have to sort of hone in on what you’re really trying to understand with the word. I think the one that’s most common is simply what we’re conscious of at any moment, what we’re aware of-
Newsome: A phenomenal awareness, like the philosophers call it. Many of your listeners will be familiar with David Chalmers and his notion of the hard problem of consciousness and the easy problem of consciousness. If you’re not, it’s definitely worth getting familiar with them. Chalmers says that there’s some things that neuroscientists are going to solve. We’re going to solve the easy problems. We’re going to solve attention, we’re going solve memory, we’re going to solve visual perception, we’re going solve visual coordination: all these features of conscious beings we’re going to solve, because we can see in principle the outlines of an answer to them, even though we’re far from having any details.
But what he says is the hard problem is why should some biological machinery hooked together in a particular way, why should there be any internal feelings at all that go along with that, that we’re conscious of? Conscious of being happier, conscious of being sadder, conscious of seeing red, or conscious of seeing green. Why is there that phenomenal experience?
And one of the things I’ve learned as a neurobiologist is that I can ask questions up to a certain point in animals, like I can electrically stimulate different parts of the brain, and I can elicit very sophisticated kinds of responses and behavioral responses, and yet, I do not know what that animal is actually feeling at the moment. There’s this first-person experience of our beings, and presumably of other animals’ beings, that is very difficult to know how we would describe that in any kind of objective terms, any kind of math that you could-
Li: The Qualia experience.
Newsome: Qualia, exactly. And that’s the hard problem. I’ll tell you, most neuroscientists, the large majority of neuroscientists, would deny that there was a hard problem of consciousness. It’s almost an ideology, honestly, because neuroscientists believe in the supremacy of their field. It’s a very deep commitment, and that once we get a mature neuroscience 500 years from now, however long it takes, there will be nothing about the brain or the mind left to explain.
And people who take the hard problem of consciousness seriously say, “It may be that an intrinsically third-person science cannot account for what is intrinsically first-person experience.” That there just may be a category of mismatch. And so I give credibility to that, but I’m an unusual neuroscientist in giving credibility to that.
Li: Yeah, you’re a very open-minded neuroscientist. I remember as a physics student at Princeton some physicists said that humanity is incapable of understanding the universe to its deepest depth because we are part of it, and it’s hard to study within something, the totality of that thing.
But just to be a little more concrete on a consciousness note for AI, one of the narrower definitions of consciousness is awareness; not even this deep awareness, but contextual awareness. And one of my favorite quotes of AI come from the 70s, that goes like this (and keep in mind, this is the 70s). It says: “The definition of today’s AI is the computer can make a perfect chess move without realizing the room is on fire.” Of course, we can change the word chess move to a different game, like Go or anything else, but today’s AI algorithm, not to mention the deeper level of awareness, does not even have that contextual awareness. And this is five decades later, so we have a long way to go.
As scientists, especially someone like you with long career in neuroscience and computational neuroscience, and someone like me who is trained both in AI and neuroscience, inevitably I think about the question: is there a higher science, or a unified framework to think about intelligence? We talk about the current gap between NI — natural intelligence — and AI. We talk about the potential continuum that we can close this gap. But maybe at the end of that is a unified science, or Newton’s laws, or the general relativity of intelligence. Do you foresee that? Do you have any conjecture of that?
Newsome: I don’t think I’m smart enough to see that far into the future, Fei-Fei. I think that’s a laudable thing.
I was a physics undergrad, so it may be physics but it was part of my formative experience, even though I never got terribly sophisticated. And I like unification, I like coherence, I like the idea that physical reality is one thing fundamentally, and that we ought to be able to understand it through successful approaches to physical reality.
I do have this question about consciousness out there that I don’t know if our third-person science will be able to deal with, but in terms of intelligence, of general intelligence, flexible intelligence, contextual sensitivity, I think we should be able to get there. I think Chalmers would say, “That’s part of the easy problem of brain function or consciousness.” And I think we should be able to get to some general principles.
I think we might not get to them until we take the AI systems and put them on the robots that actually have to make their way in this world in order to survive. Only then will the robot care whether the room is on fire.
Li: Yeah, actually I really want to accentuate your point because you said multiple times you believe in a physical body, and I totally agree with you. You evoked Descartes’ “I think therefore I am,” which removed the physical body but just looking at evolution as well as today’s working AI and robotics and machine learning, I also think that the embodiedness of agency is critical, at least in the development of human intelligence, and will become more and more relevant and critical in artificial intelligence. So I agree with you.
But I’m still hoping the Wu Tsai Institute or HAI, hopefully together, will host a new generation of scientists that one day will give us the universal law of intelligence and unify some of these questions in intelligence, don’t you think?
Newsome: Yeah, Fei-Fei, it’s interesting. I’m not quite that ambitious yet. In just thinking about fundamental principles of computation of the brain, and information processing, and information extraction, and organization, and decision making, and memory, and learning, I don’t even know that there are going to be a single set of principles for the brain. I think that different brain structures have very, very different architecture. The cortex is very different from the basal ganglia. The basal ganglia are very different from the spinal cord. The spinal cord is very different from the hippocampus. And all of those are very different from the cerebellum. And I think that the computational principles, even though they all work with action potentials, they all work with neurotransmitters, they’re all subject to neuromodulation, there are certain things like that, that are universal probably across the brain.
But the rules of computation and algorithms that are instantiated in those neurons of very different structures may simply be different. It may be that there’s going to be one theory and one set of principles for the cerebellum, and a different one for the cerebral cortex. So I don’t know that there’ll be any general unified theory of mammalian brain function. It may be a collection of allied theories. And then you’ve got to have theories about how those circuits interact with each other, or produce behavior.
But I think there’s so much headroom for progress, so much new data being acquired and I’m optimistic. I think this is a great time to be a neuroscientist. It’s obviously a great time to be an artificial intelligence person, and I hope that these two great human enterprises do converge in a meaningful sense at some point.
Li: I think that is such a great note to end on, and just to echo your call for action, especially to the students out there, that it’s a phenomenal time to be considering these two fields, and especially the intersection of the two fields. I think some great discoveries and innovation will come about because of this intertwined joint adventure between neuroscience and AI.
So, thank you so much, Bill. We’ve often said that you and I can sit here and just talk for hours and hours, and every time I talk to you, it’s just so inspiring and just so humbled that I’m your colleague, and working on this together.
And thank you to our audience for joining us. You can always go to the HAI website to visit us, or go to our YouTube channel to listen to other great discussions with leading AI and thought leaders around the world. Thank you.
Stanford HAI's mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.