The Wonder of the Emergent Mind (with Gaurav Suri)

Gaurav Suri: Thank you, Russ. Pleasure to be here.

Gaurav Suri: Wonderful, thank you.

Gaurav Suri: Right. The experience of living is that we do things. We wake up, we go to the mall, we choose a taco instead of a burrito, we perceive things, we understand things. The question is: how does this happen? And a really inviting answer is that it happens because we have intelligence built in to the components of our mind: that there are neurons that know how to understand, that there are neurons that know how to speak a language or perceive.

When you actually look at neurons, they don’t appear to be doing any of these things. All they do is they activate and maybe they connect with other neurons.

So, the idea of the title, and really of the book, is that intelligence, our intelligence, but also intelligence of our machines, emerges from the interaction of simple processing units that themselves are not intelligent in the way we understand intelligence, that they themselves are doing something very simple, and yet the whole has properties that none of these neurons have.

So, Russ, this is the definition of emergence: Emergence is a property that’s present in the whole system, but not in the component parts in the same way. And, the thesis of the book is that intelligence in our brains and in our machines emerges in that it’s not present in the units–the neurons or the neuron-like computing entities–but is present in the whole. And, emergence, as you know, is not a rare thing in the universe, that it applies to brains. Emergence is everywhere, from galaxies to water, to ants. It exists everywhere, and the book makes the case that that is how we should approach our pursuit of understanding our intelligence.

Gaurav Suri: I think this is a really interesting example and a seminal example, and I am delighted to lead our conversation with this. Because I, after this example, Russ, began to think of our brains as a colony of ants. That’s a metaphor, but I think a good metaphor for what’s going on.

So, I grew up in India and I grew up in an era where there was no social media and bad TV. And so you were outside. And one of the things that I did when I was outside was look around in the world, and one of the things I saw were ants. This example that you were talking about consists of ants going from their nest to a source of food and coming back. And, the line of ants can become pretty dense. So, they’re going from their nest, they’re getting the food, and they’re coming back maybe with a particle of food.

And, as a young person, what I did was I placed an obstacle on that train. So, if you imagine a band of ants, you put an obstacle there and you see what the ants do.

Well, you do something interesting, which is you make the obstacle have a short way and a long way. And initially–it’s quite dramatic to actually see this happen–initially, the ants go 50-50, right? So, 50 going the long way, 50 going the short way. And in a few minutes something amazing happens. Nearly all the ants, not all, but nearly all the ants are going the short way. And, as you nicely ask, ‘Well, what’s going on?’

And, when my son was young, I would ask him this question and he’d say, ‘Well, maybe the queen ant is communicating to the other ants,’ and well, first of all, they’re not communicating in that way, and second of all, they don’t have a conception of distance. Right?

Ants by themselves do not distinguish themselves by their intelligence. You put an ant on a flat surface and the ant is pretty much going to walk around until it dies of exhaustion. And yet, together they’re capable of extreme intelligence, like building cities and finding the short way around obstacles.

So the question is: What are they doing? Well, they’re doing something very simple. Just like we talked about neurons doing something very simple, which was activate and connect, ants do something very simple, which is secrete pheromone and follow the pheromone trail. That’s what they do.

Now, it turns out that if there’s a long way and a short way and the ants are 50-50 and one ant goes the long way and one ant goes the short way, the ant that goes the short way is going to reach the destination first. It’s going to have a choice, ‘Gee, which way should I go back?’ Well, it’s going to go back the way it came because it laid the pheromone while coming, and so it’s got more pheromone on the short path than on the long path, so it follows the short path.

A little bit later, the ant from the long path comes and it has to decide, ‘Which way am I going to go?’ Well, it has pheromone on its path, but now there are two trails of pheromone on the short path, the first ant coming and going. So, it comes back the short way, and boom, that’s all it takes. It’s like there’s a greater pheromone concentration on the short path, and that leads to the phenomenon.

Gaurav Suri: Right. There’s a hundred billion neurons, give or take. The latest number is 86–I think that’s subject to revision–86 billion neurons. But, let’s–

Gaurav Suri: I am not at all sure about that, and I’m also not sure that counting neurons is a good way to count intelligence. But, Russ, what you’re saying is fascinating because our brain has electricity in it. This is not like a line from a Frankenstein sort of story. Our brain is composed of these neurons. There’s other things in the brain, but the principle information processing engine of the brain are these neurons.

And, a neuron–think of a neuron like it’s sort of like a tree. So, up top there are these dendritic branches. So, imagine the massive branching happening, and then there’s a long trunk with some shallow roots. The long trunk is called the axon, and, like, a tree ends in the ground, but a neuron–imagine the axon of a neuron, the tail of a neuron opening into the branching–the dendrites–of another neuron.

So, there are these structures that are in close proximity to each other. And neurons, at a high level of abstraction, do two things. They generate bursts of electricity called action potentials. These are actual bursts of electricity. So, you can hook them up to a device that transforms electrical signals to audio, and you can hear bing, bing, bing, pshh. It sounds like that. And, that’s the sound of a neuron having electricity. It’s quite marvelous.

So, that’s the one thing they do; and the other thing they do is they connect with each other. They may be connected from birth. So, we are born with pathways of neurons that lead from our retina to our visual cortex in the brain, and in fact, that is how it was first discovered–this is a slight tangent, just 30 seconds. It was first discovered that it is the brain that is the seat of intelligence because a gladiator or someone died and they followed his optic nerve. And, some people thought–I think some Greek philosophers thought–that it is the heart that is the seat of cognition. But, you follow the optic nerve to the train–the axons that are strung together–that make a pathway, the gray matter track, and that goes to the visual cortex. That’s in the back of the brain.

So, these connections are innate, meaning some of them exist from birth; but they’re also modified by experience.

So, when we learn something new–like, I learned something when you were talking about the Blue Angels just now, and that was a nice analogy for me because oh yeah, murmuration and communication and emergence, and here the emergence is maybe the birds looking at each other, and there the emergence is maybe the pilots to each other, right? But it is an emerging system. So, I learned that. And, when we learn some things, we represent knowledge in our brain by making new connections. Neurons that weren’t connected before are getting connected.

And that’s the only two things that neurons are doing. And from these two things, the emergence, the intelligence, is emerging.

Gaurav Suri: Yeah, that’s a lovely, beautifully posed question. Thank you for it.

Let’s take a simple example. This experiment that I’m going to describe was done in the 1970s by two, I think, brilliant psychologists, Nisbett and Wilson. What they did was they presented four stockings–let’s label them A, B, C, D–to people. And, they asked them to choose whichever stocking they like, and people chose the fourth stocking in general, and when asked, ‘Why did you pick this stocking?’ some people said, ‘I like the thread count. The hue of the color is slightly different.’ They all had their reasons. Meantime, all the stockings were identical, and it’s a well-known fact that when you give people a list of identical things, they’ll most likely pick the last thing. That’s just a choice pattern that people have. And so, nobody said, ‘I picked it because this was the last thing in the order, and these were identical.’ They had a reason.

And, there are some very beautiful experiments that–so there are two hemispheres in our brain, the left and the right, and they’re connected by a bundle of nerves called the corpus callosum, and for some purposes, a while back, patients, they would cut off the corpus callosum to prevent fits and whatnot. So, now you have two hemispheres that are not talking with each other, and it’s possible to show the visual field of the right hemisphere, which, for many people, does not dominate in language–show them a message like: ‘Start crawling on the floor.’ And they can’t process that they’re crawling on the floor because you showed a message, because the corpus callosum is cut off and that message hasn’t traveled to the language center, that hasn’t made awareness, and whatnot.

And, they come up with a reason. So, they will say things like, ‘I think I dropped my keys,’ or ‘go to the bathroom.’ ‘Why are you going to the bathroom? You just went.’ ‘I think I need to wash my hands.’ So, the thesis of the book is that this system of associative neurons that are interacting with each other, rather like ants interact with each other, are also producing our justification and stories for why we do the things that we do. It’s the same system that’s producing the choice is also producing the reason for the choice. And sometimes the reason has a lot to do with the choice. ‘Johnny, why don’t you go and play in your friend’s house?’ ‘Oh, last time I was there, he beat me up.’ That reason probably has a lot to do with the underlying neural net, also known as the brain.

But, the astounding thing is that there are different systems, and they may not–and we humans are master storytellers, and we often tell stories that put us in a good light and make us virtuous and heroic. And, these are accounts; these are stories produced by the underlying network.

So, let’s take a very simple example. Let’s say you go to the refrigerator for snack, and somebody says, ‘Russ, why’d you go to the refrigerator?’ ‘Oh,’ you might say, ‘I’m hungry.’ Or, ‘I’m thirsty.’ And, true, that’s true. What’s happening underneath? Well, you have thirst-detecting neurons in your brain. The way these neurons work is they detect the salinity level in your blood, and if the salt level is too high, that means you don’t have enough water. And they start to activate, right? Bing, bing, bing, bing. Or hunger neurons might start to activate. And they’re connected with experience. They’re connected to another network that initiates signals to the muscles to move in the direction of the refrigerator.

So, these neurons that are interacting are you go to the refrigerator. They don’t necessarily have to do anything with the linguistic conscious feeling ‘I’m hungry’ or ‘I’m thirsty and I’m going to the refrigerator.’ You might have that feeling, or you may go to the refrigerator out of habit because you usually go at 6:00 when you come back from work. The point is that these systems do their thing. They make us move, they make us do things in the world, they make us choose–and we’ll talk about economists and pleasure and pain in a minute. But, they are busy interacting with each other, kind of like ants, and they are producing our thought and our actions.

Now, is that disturbing? Well, I’d say no. And, I want to mark a reason for hoping we can come back to that later in the conversation.

Gaurav Suri: Well, that’s a great example. I love that example. I think this–it’s not quite a habit. So, a habit is something that you almost–it’s more automatic. I call it action readiness. So, the way I describe it, Russ, is that you have a greater action readiness to snack when you’re eating–

Gaurav Suri: When you’re reading, sorry.

When do we get action readiness? Well, watching somebody would do it. A particular context would do it. Lots of people eat popcorn while watching a movie. Even if they don’t want popcorn, even if the popcorn is stale, they will eat it, right? Now, interestingly, if you offer people stale popcorn in a conference room, they won’t eat it, but in a movie house, they’ll eat it.

Gaurav Suri: So, here’s the amazing thing–this struck me when I was doing my Ph.D. So, a little bit of a biography, because I think this is important context. I worked in industry for many years, I was a management consultant, I was a partner at Deloitte Consulting, and I saw people deciding in boardrooms a lot, right? And I would–and these are big decisions: What is the price of our inkjet printer in South Korea? And I’m thinking, was this really decided based on the data that I’m presenting, or was it because this person showed up to the meeting and that person didn’t, and there’s light streaming in, and so-and-so wants to go to the bathroom? There are this conspiracy of interacting things: Just like neurons are interacting, people are interacting. And I was really struck by people have patterns, people have greater action readiness to do certain things.

So, I come, in midlife–consulting was kind to me–and in midlife, I had the luxury of coming back to academics, and I did my Ph.D. And one of the first experiments I did–this is crazy, I still can’t believe this experiment–if you show people two pictures, one of a beautiful scene in nature, and one of, say, a horrific mutilation, and ask them to pick one picture, 90%-plus of them will pick the beautiful scene in nature. Not a surprise. A little bit of a surprise that there’s still about 10% that want to see the other picture, but 90% is close enough, as close as you get to 100% in human behavior.

But, if you change the experiment slightly, what you do is you show them the horrific picture and teach them that if they press the S key–S for switch–they can switch to the beautiful nature picture. It’s the same choice, essentially, right? It’s just that now, a proactive action is needed. It isn’t framed in the context of choice. Now, in the case of proactive action, only 50% switch from the bad image to the good image.

But here’s the amazing thing. If you, instead of S, make it a forward slash key–which people don’t usually press–now, instead of 50%, only 30% switch.

And here’s the–this is the experiment I did, which is if you pretend them[?] to copy the HTTP line of code, which has a bunch of forward slashes–so they’re copying code. They don’t know why they’re copying code, but they’re copying code, so they’re writing a bunch of forward slashes. Now, if you ask them to do it, that 30% goes back to 50%, because they pressed slash a bunch on the previous task.

So, things that we do, things that we’re used to doing, profoundly influence the things that we do. The things that we do are influenced by preference, yes; but they’re also influenced by what we’re looking at, what our attention is directed to, what our action readiness is, what we saw somebody else doing, what the incidental associations are.

Gaurav Suri: Yeah, that’s great. So, one metaphor that my co-author, Jay McClelland, recounts, is to start thinking of neural networks as reservoirs of interconnected pools of water. Like, you know how in a stream, how water falls: there’s little pools, and there’s a channel from one pool to the other pool. So, think of these little micro pools as units, which are populations of neurons, and think of the channels between them as the connections between these neurons. And, the more water that goes from one pool to the other pool, the deeper the channel. Right? And, the thought here is that the more a neuron is co-occurring with another neuron–so if I say green and you say grass, what happened? Well, what happened is that neurons corresponding to the word green have a deep channel or a connection with neurons connected with saying the word grass or thinking of grass.

And, what made this channel? It’s the frequency of exposure.

So, this was Canadian neuroscientist Donald Hebb, came up with this idea that if two things are co-active, they become connected. So, neurons that fire together wire together. It’s a simplification of Hebb’s rule. And it’s a beautiful rule. I’ll repeat it: Neurons that fire together–meaning they are co-activated together–wire together.

So, imagine your brain: We’ve been imagining your brain as an ant colony, and now imagine it as the system of pools of water in a stream, maybe in a waterfall, and the pools fill up. Some pools fill up, depending on how the water is falling. Water comes and goes, but the traces of the traffic between the pools, that stays; and that’s the knowledge of the system. Right?

So, this is a startling idea, but it’s useful to think of your thoughts as patterns of electricity in your brain, patterns of activation. Thoughts come and go, just like water comes and goes in a stream, but your knowledge are the channels, are the connections, between neurons. And, this is why people respond differently to the same input, because they have different connections. They have different channels that takes the input that they’re seeing or smelling or tasting and sends it to different parts of the brain based on, in part, their experience. And, if you’ve had different experiences than mine, then your channels are going to be different, your pathways in your brain are going to be different, and you’ll respond differently to the same input than I will.

Now, what makes these channels? Well, the point here, the central idea, is that channels can get made just by repetition. Eating while you’re reading, do it a few times, and you’ve got that action readiness, you’ve got that channel.

Gaurav Suri: Right. So, let’s talk about free will, because it’s come up, and it’s centrally related to what we’re talking about here. The picture that we’re painting about these channels shaped by experience is a picture of a deterministic system–

Gaurav Suri: And, ‘deterministic,’ what does that word mean? The deterministic system means that a system is–operations of a system are predictable based on its current state and its input. Right?

Now, well, we learn from experience, so we’re changing our brain, and so we’re not going to respond the same way to the same input. Literally, if I learn to juggle, I’ve got so many new connections in the brain that you can see them. You can actually see white matter tracts that didn’t exist before. So, the brain is plastic. It’s changing. But, the thing is, it’s a deterministic system, and yes, absolutely, the case we’re making in The Emergent Mind is that it’s a deterministic system.

But there’s a big ‘but,’ and this is why I’m really hesitant to say free will is not real. There’s three ‘buts,’ so bear with me.

The first ‘but’ is that free will is a very useful construct. It’s a construct, because if we don’t have free will and we don’t hold people accountable for what they do, there will be chaos. So, even if the system itself is deterministic, the construct is very useful, and we need that construct, we need to live by that construct. That’s Point Number One.

Point Number Two is that for all practical purposes, even though it’s a deterministic system, it is such a complex system that it’s computationally pretty close to not being easily tractable. It’s an extremely complex system, right? You can’t tell whether something’s going to land heads or tails, even though that’s a deterministic system, depending on the force of your–

Gaurav Suri: There’s nothing magical.

The third thing is that we have goals; and goals, too, are deterministic. It’s predictable where these goals come from. Goals are basically, I think it’s very useful to think of them as memory structures that influence action. So they’re a sub-network. And, how do goals come? We will talk about that later. But, goals are essentially sub-networks of activation that are influencing what you do at a current state.

Now, goals–when people talk about free will, they really are talking about the pursuit of goals. And goals live in a deterministic system. And, I think it’s really a disservice for a neuroscientist to say that, ‘Yeah, we live in a deterministic system, and free will is not real, and you guys are full of it,’ because free will is a very nuanced thing. It’s a useful thing. It reflects on goals. It’s talking about things that are complex. So, I don’t think it’s useful to go around saying that, ‘Look, free will isn’t real.’ I think it’s a useful construct. But, the point is that our goal pursuit is coming in a deterministic context.

Now, we are a process, Russ, absolutely–you said like an animal running around or like water flowing downstream–we’re a deterministic process. And, I think that we are a part of the universe, no more than the trees or the stars, but also no less. We’re a process. And I find great comfort in that. My understanding of myself–I’ve stopped thinking of so-and-so as just fundamentally evil or so-and-so as–like, we are processes. We’re processes shaped by our context, and we get to have goals based on our experience and our bodies and the demands, and we follow those goals, and we watch ourselves unfold in this life, we see ourselves unfolding, and that’s pretty magical.

Gaurav Suri: Well, yeah, I could see how one would think it’s troubling if one thinks of this internal spirit that is the source of one’s actions.

Gaurav Suri: Yeah. You, Russ, you are an emergent consequence of the little parts of biology that are made; and I think there is a lot of beauty in watching yourself unfold.

Now, by the way, there are things that we do that seem effortful to us–so just paying attention–that feel an awful lot like an inner us is exerting control.

So, there’s this famous task in psychology; it’s called the Stroop task, and the Stroop task is in one condition, you’re supposed to read the word, say green, and you say, ‘That’s green, that’s black, that’s white,’ whatever. But, in the other case, you’re supposed to ignore the word, which is a color word, and instead identify the ink that the word is written in. So, if the word green is written in red, you’re supposed to say red, not green.

Now, it turns out–this is reliable; it’s been done thousands and thousands of times–and reliably it’s true that in that color-naming task, people are slower than in the word-reading task, because for all of us, word reading is much more automatic than color naming.

But, what’s happening when we’re naming the color? This is the question. It’s fundamentally an attention task. Now, what does it mean to pay attention? It means to pay attention is to increase the activation in a particular stream. And, again, this is a deterministic thing. It’s guided by your goals; your goals of doing this task is guiding it. But, you have the capacity to pay attention and you have the feeling of effort. So, this is where this ‘me,’ this thing you’re talking about, this self-is coming, because you, the self, are paying attention and pursuing your goals, and that’s what’s giving you this notion of achievement. And, the point is that that’s valuable and beautiful and no less beautiful than it’s deterministic.

Does the number three exist in nature? What is the concept of three? So, when we are children, your mom says, ‘That’s three ducks,’ or, ‘That’s three trees,’ or, ‘That’s three dogs.’ And, we form a concept that is the intersection of all these experiences we’ve had with three things, and this concept emerges of three. Three is a concept. So, this effortful goal pursuit is a concept that emerges, again deterministically, but it’s a concept. And, just because it’s a concept, doesn’t mean it’s less real. Is three less real to us? Where is three? Show me the platonic idea of three in nature. No, it’s a concept; and yet, it feels as real and we honor it and we use it. Free will is a concept that emerges in this deterministic system, so let’s honor it and use it.

Gaurav Suri: Yeah, yeah, lots of ideas. This is what I think about all the time, Russ. So, yes, I’m a neural network guy, and I’m going to think about this in neural network terms.

Now, first, you’re laying in bed and it’s a stream of consciousness, and one thought is connecting and cascading to the other, and a few things are happening. Maybe light is coming in through the blinds and that’s triggering the day is started. Maybe you’re hearing the birds outside, the day has started. And, with the day started, a common action-ready thought is, what are my goals for today? Now, somebody else might not get that thought. You, Russ, from your life, are in a place where maybe you get up and think that thought, which is, what are my goals for today? And, that thought, like any other thought, is a pattern of electricity, it’s a pattern of activation in your brain.

Okay. Now, when that pattern happens, it’s activating potential goals. Right? And, you remember that you’re going to talk to me, and you finished reading the book, and you’re framing your questions. That, what are my goals now stream of consciousness, like, still is coming to: what are my goals for today? And, now all of a sudden it’s, like, ‘Oh yeah, yeah, I need to ask that guy about Adam Smith.’ I’m making this up.

Gaurav Suri: No, no, the quote is not a caricature. I’ll talk about that in a second. I mean, the quote is Adam Smith–I think the quote is him reflecting on how powerful body the [?]–

Gaurav Suri: Okay. So, now you have this goal, and this goal is: you want to ask me a specific question. You want to be precise about that question. So, what’s happening is, this goal is directing your attention to that particular question, and you’ve focused on that question. Still a very deterministic system. By the way, the act of giving attention has different experiential qualities, right?

Gaurav Suri: And, Daniel Kahneman talked a lot about that. And, Daniel Kahneman’s answer was that, ‘Well, it’s two different systems.’ I said, ‘No, no, it’s the same system.’ It’s the same system, but different kinds of things are happening in the same system. In this system, you are influenced by the goal, which was activated not through magic, not through the inner self, some magical inner self, but through a deterministic flow of activation. And, now that goal is parsing your activation to really zone in on that particular question where there’s ambiguity, and you are examining those chains of thought, and that feels different. And, all of a sudden you are in this different mode experientially without the underlying pattern of connectivity, the pattern of neural network, changing. It’s the same neural network. Because, what this does for us that the inner spirit doesn’t do, is it explains how we might arise from the things that we see in our brain.

The option is to be a dualist. So, there’s a story in the book where Descartes–the French philosopher, mathematician–he’s walking around in a garden, and he steps on a stone, and all of a sudden the arm of a nearby statue moves. And Descartes is, like, ‘Wait, what happened here? How did the arm move?’ So he talks to the gardener, and the gardener explains–in French, I’m sure–that there are hydraulic pipes that go from the stone to the statue, and that when he steps on it, the water pressure through the pipe makes the arm move. And, many of us would have said, ‘Oh, okay.’ But Descartes, he’s struck by this. He’s, like, ‘Ah, maybe this is the mechanism by which we move our hand away from a fire: that maybe there’s some signal that the fire, like a water, transmits to our brain, and that the brain then sends a signal back into the pipe, and it’s all done with water-like pressure.’ And, he imagines the nervous system to have microtubules filled with fluid.

And Descartes is wrong in every way. But he’s brilliantly wrong, because now he’s after a mechanistic way of explaining the things we do. And he makes a lot of progress. So he says, ‘I can imagine this machine’; but he stops short because he says, ‘When we do mathematics or fall in love or write poetry, it can’t be these tubes.’ He can’t imagine these tubes.

Well, with a neural network, you can. And you not only can imagine them, you can give useful accounts of how these things emerge from the interactions of simple processing units. And, Descartes’s answer was, ‘Yeah, moving away from a fire mechanistic–animals have this mechanistic stuff, but we have this inner spirit that lives in our pineal gland.’ And so, that’s how dualism started.

And, in some ways, that tradition, Descartes was making progress. But he didn’t go far enough. He didn’t make the possibility that perhaps it’s useful to think of our entire brain or all our actions coming from mechanistic processes, like water in the pipes that he was thinking about. And so, he is left with ‘the spirit.’ And people have tried to weigh dead bodies after–the spirit is not that useful because if the spirit knows that you want a taco instead of a burrito, well, how does the spirit know that? How does the spirit know that you like this and not that? It’s not explaining anything, right? I mean it’s basically a restatement of the problem.

Gaurav Suri: Well, let me stop you there.

Gaurav Suri: The neural network is not the same as the soul because what is possible–

Gaurav Suri: But, it’s more helpful because what it shows is that it can–with the same inputs, with architectures of neural networks, you can see how the output emerges. You can see, like you can see a ball rolling down in your example earlier. You can trace the activations and you can say, ‘Oh, this is why false memories happen, or context happen. You can trace these activations; and it’s magical. Like, when you see this emergence happen, and you trace it, and you understand operations of the mind like you understand a garage door opener, for me, this is transformative. And so, while the soul might be a black box, the neural network isn’t. You can trace it. [More to come, 56:54]