"The Emergent Mind: How Intelligence Arises in People and Machines" | Jay McClelland

The AI revolution of the past few years is built on brain-inspired neural network models originally developed to study our own minds. The question is, what should we make of the fact that our own rich mental lives are built on the same foundations as the seemingly soulless chat-bots we now interact with on a daily basis?

Our guest this week is Stanford cognitive scientist Jay McClelland, who has been a leading figure in this field since the 1980s, when he developed some of the first of these artificial neural network models. Now McClelland has a new book, co-authored with SF State University computational neuroscientist Gaurav Suri, called "The Emergent Mind: How Intelligence Arises in People and Machines." 

We spoke with McClelland about the entangled history of neuroscience and AI, and whether the theory of the emergent mind described in the book can help us better understand ourselves and our relationship with the technology we've created.

Learn More 

New book sheds light on human and machine intelligence | Stanford Report

How Intelligence – Both Human and Artificial – Happens | KQED Forum 

From Brain to Machine: The Unexpected Journey of Neural Networks | Stanford HAI

Wu Tsai Neuro's Center for Mind, Brain, Computation and Technology

McClelland, J. L. & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. [PDF]

Rumelhart, D. E., McClelland, J. L., & the PDP research group. (1986). Parallel distributed processing: Explorations in the microstructure of cognition. Volumes I & II. Cambridge, MA: MIT Press.

McClelland, J. L. & Rogers, T. T. (2003). The parallel distributed processing approach to semantic cognition. Nature Reviews Neuroscience, 4, 310-322. [PDF]

McClelland, J. L., Hill, F., Rudolph, M., Baldridge, J., & Schuetze, H. (2020). Placing language in and integrated understanding system: Next steps toward human-level performance in neural language models. Proceedings of the National Academy of Sciences, 117(42), 25966-25974. [

Thanks for listening! If you're enjoying our show, please take a moment to give us a review on your podcast app of choice and share this episode with your friends. That's how we grow as a show and bring the stories of the frontiers of neuroscience to a wider audience.

We want to hear from your neurons! Email us at at neuronspodcast@stanford.edu

Learn more about the Wu Tsai Neurosciences Institute at Stanford and follow us on Twitter, Facebook, and LinkedIn.

Transcript

Nicholas Weiler (00:07):

This is From Our Neurons to Yours, a podcast from the Wu Tsai Neurosciences Institute at Stanford University, bringing you to the frontiers of brain science.

(00:24):

Today on the show, the Entangled History of Neuroscience and AI. In the eighties and nineties, a new paradigm for understanding the mind arose in cognitive science. This approach proposed that patterns of activity distributed across vast interconnected networks of neurons could self-organize based on our experience, not on any explicit rules or programming, to allow us to make sense of the world and pursue our goals.

(00:54):

Under this paradigm, the mind is an emergent phenomenon, and it's worth saying another word about that idea of emergence. Throughout science there are complex phenomena that only exist when many smaller, simpler pieces come together. For example, water is wet, but water molecules are not. Ant colonies are pretty smart, but individual ants are pretty dumb.

(01:20):

One of the key ideas of neural network theory was to show that complex mental states like perception, memory, and decision-making could emerge from simpler patterns of interactions between interconnected neurons, even highly simplified artificial neural networks in a computer model.

(01:38):

This was a totally different way of conceiving of the mind, and this approach has been incredibly influential in neuroscience. It's led, for example, to the sort of brain-machine interface breakthroughs we've covered on the show before. But of course, the neural network approach to thinking about the mind also gave direct rise to the revolution we are currently experiencing in AI. Today's large language models are all based on artificial neural network architectures that were originally developed to study our own brains.

(02:08):

So the big question is this. If our minds and modern AI are built on the same principles, what does that mean for our understanding of ourselves and of the future of these technologies?

(02:21):

Today's guest is cognitive scientist, Jay McClelland, a professor of psychology who leads the Center for Mind Brain Computation and Technology here at the Wu Tsai Neurosciences Institute. Jay has been a leading figure in this field since the 1980s when he developed some of the first neural network models. Jay has a new book, co-authored with his collaborator, Gourav Suri at SF State University called The Emergent Mind: How Intelligence Arises in People and Machines. I have to say, I've been thinking about it ever since I read it.

(02:53):

Obviously, there's too much in a book like this to cover in one episode, and I'm not sure I could do it justice in a longer conversation, but this is a topic we'll surely come back to many times on the show. Because the question of what the mind is used to be philosophical, but with the AI revolution, it's becoming increasingly urgent.

(03:18):

At the simplest level, what is a neural network?

Jay McClelland (03:21):

So a neural network is a model of the brain. It's a network of interconnected elements, each of which has properties similar to the properties of actual neurons. There's a necessary abstraction whenever you build a model, it needs to be a simplified representation of the actual true complexity of real neurons to give insights, but it's still intended to capture what the modeler considers to be some of the important features of the real neural networks in our brains.

(03:57):

A neuron is a cell, and you can think of it like a computing element. And in fact, there was a moment, almost a hundred years ago now, when people began to see, "Oh, a neuron could perform a logical computation." A neuron can receive signals from many other neurons, and it sends out signals to other neurons. Just like a logic gate in a circuit, a logic gate that sends out a signal when two other signals are both present is called an AND-gate. A logic gate that sends out a signal when either of two other signals are both present is called an OR-gate.

(04:35):

A neuron is like a logic gate except that it's getting up to a hundred thousand inputs. So we have to think in terms of the aggregate influence of all of the factors that impinge upon it. And then secondly, its output is graded and somewhat noisy reflection of its inputs.

Nicholas Weiler (04:57):

I sometimes think of this as a massive game of telephone, where in telephone one person hears a message from someone else and then passes it along and things get changed a little bit along the way. Except in this situation, each player is hearing messages from hundreds of thousands of different inputs. And then based on all of that, whether they're getting messages at the same time, how strong those messages are, whether they're aligned with the thing that neuron is attuned to, it goes ahead and sends its message to tens of thousands of other cells.

(05:31):

So it's sort of this flowing of activation, which is a word you use a lot in the book, that moves through these systems activating cells when the input meets some certain criteria in order to pass the message along. And if it doesn't meet the criteria, the message would stop there.

Jay McClelland (05:48):

Yeah, I like that way of thinking a lot. A fundamental part of my thinking is that no one neuron ever really makes a difference. Our brains consist of like 86 billion neurons. So it's like something that depends on the population as a whole, having capacities which the neurons are required to embody but do not themselves really have as such, as individuals.

Nicholas Weiler (06:19):

Well, you had this beautiful image towards the beginning of the book of a waterfall with the water sort of cascading through multiple channels into these system of pools. And sort of thinking of the mind through these networks of neurons in a similar way, which is no one stream of water makes the pattern or makes the pools. The water will flow through millions of different small channels depending on how finely you look, but it's the system, it's the pattern, and it's the way gravity pulls the water through the system that really matters.

(06:52):

So in the 1970s and eighties, I know that the concept of neural networks had actually originally come up in, as early as the forties and fifties. But when you were getting involved with this and your colleagues were getting involved with this in cognitive science, trying to model aspects of how the brain works by creating simplified models of small networks of neurons on computers, what did those models look like? Just sort of the fundamental of what is a simplified neural network model look like, and maybe, if this makes sense, how was that different from the way that people were thinking about information processing and cognition in the brain at the time?

Jay McClelland (07:30):

Great question. Let me answer those two parts separately. So we began with the idea that we would try to capture cognitive processes using very simplified neuron like processing units, and we didn't want them to correspond to individual neurons, but to have properties that neurons have. Specifically, each one of these individual units was capable of receiving inputs from many other units by way of connections, which had a graded strength. So by a graded strength, I mean a real value strength, which could either be positive or excitatory, or it could be negative or inhibitory. So that when an individual unit is receiving inputs, it is kind of adding all of these influences together, and their net effect is what determines its internal degree of activation.

Nicholas Weiler (08:29):

So it's sort of summing up its inputs. Some of them influence it to become more active, some less active.

Jay McClelland (08:34):

Yeah. We actually imagined that these units were trapped within a range of activation values. The idea was that the net input they received was unbalanced excitatory, it would push them towards the maximum. If it was unbalanced inhibitory, it would push them towards the minimum. In the absence of any input, they would tend to settle back to a resting level below zero.

(08:57):

And when they're below zero, they're not actually sending a signal out to other units. They're kind of just quiet, and it's only when they get to zero that their activation begins to propagate out. So they only send positive signals. Like neurons, they send action potentials, which are basically a degree of excitation-like signal. So it's like a neuron in that respect.

Nicholas Weiler (09:20):

And in neurons we often count the activation as how quickly is the signal coming out? Are they sending action potentials once a second, 10 times a second, a hundred times a second, even though a neuron only is on or off, you can get these sorts of graded levels of activation.

(09:38):

So you've got these artificial simplified neurons that are taking in inputs and sending out signals to other neurons. And then there's also a component where in these networks they can sort change how much they influence each other. The connections between the neurons are plastic, as we say in neuroscience, to represent the fact that these networks can change how they are passing information between them.

Jay McClelland (10:03):

Yeah. So there's some very basic principles that I rely on. The first is that whenever we're actually having an experience, we're representing something in our minds. It's because there's a pattern of activation over neurons within and across multiple brain areas that are supporting this. So my perceptual experience when I look out at my room or look at you is rich. It includes all kinds of features. It also includes thoughts at many levels of description, like I'm talking to a person who I know who knows something about neuroscience while I'm also seeing the expression on your face.

(10:44):

And all of this stuff is represented within my brain in terms of patterns in different parts of my brain. So my immediate experience, the underlying support of it is these patterns of activation.

(10:57):

What makes it so that the input gives rise to these patterns is the connections among the neurons. Many of those connections may be partially determined by sort of a genetic blueprint that establishes patterns of connection between different brain areas on a sort of a relatively coarse initial scale, that then subject to refinement and experience-dependent processes that change and shape those connections.

(11:26):

So activity depends on connections, but also shapes connections. One of the crucial foundational insights of Donald Hebb's book called The Organization of Behavior-

Nicholas Weiler (11:41):

Right, this is very fundamental [inaudible 00:11:43] neuroscience.

Jay McClelland (11:43):

Very fundamental idea, is that the connections among neurons are shaped by these activity-dependent processes in the sender and the receiver. So when one neuron takes part in firing another, it shapes the connection.

Nicholas Weiler (11:57):

So basically in modeling this, you're saying, "Okay, we know that neurons take in inputs from other cells." Based on the pattern of input, whether it's sufficient, they'll send their own output to a bunch of other cells, and these connections based on that flow of activity. You can model these with these artificial simplified neurons that I think you typically do it in a couple of layers where you've got an input that's getting some signal, say an image or something like that, and an output that's classifying it. Or you might have an intermediate layer of neurons that's doing some intermediate processing, and we'll get back to that in just a moment.

(12:35):

I want to get back to the second part of that question, which is when you and your colleagues started making these neural network models, was that a response to another perspective that was in the field that felt like it wasn't sufficient to understand what was happening in our cognitive processes or in the ways that neurons were talking to each other?

Jay McClelland (12:52):

Yeah, it was definitely the prevailing viewpoint was that you could think of the process that takes place between the presentation of external input, maybe an image to the eye or a printed word, and the emission of some sort of overt response by a human subject in an experiment, as broken down into a sequence of discrete stages, each of which occupied a fraction of the total time between the presentation of the stimulus and the response. So this was called the Discrete Stage Theory, and you could say, okay, well, it took 750 milliseconds for the person to press the button saying, "Yes, I recognize this stimulus from the time it was presented to the time they pressed the button."

(13:40):

And you would try to break that down into these discrete stages and then try to figure out how different factors influence the duration of each of these stages. And so there would be encoding comparison decision and response stages, and they were strictly sequential.

Nicholas Weiler (13:56):

Which is interesting because that's how you would historically program a computer to do this. It's very logical, very sequential, very patterned.

Jay McClelland (14:03):

Yeah. And the field of cognitive psychology that I entered as a first-year PhD student, my undergraduate experience was in paradigms that existed before the cognitive revolution. But the idea was that the computer provided a basis for making concrete theories of what the sequence of processes actually were.

(14:29):

The computer also provided an existence proof of something that went through a sequence of steps when it processed information, motivating people to feel that it was okay to imagine, as behaviorists had claimed it was not, that there was something going on in the head. So I think that we kind of decided that the computer was the right way to think about the mind because it had certain useful affordances for people breaking away from behaviorism. But it then led to a commitment to this discrete sequential stepwise logic-based attitude and way of thinking about the mind.

(15:12):

So I want to really speak to your particular question about what it was about that that was frustrating. And it was the fact that it just didn't seem particularly to my colleague David Rumelhart, who was a senior colleague of mine at UCSD in 1974, that our ability to figure out the meaning of and interpret what our senses provide us as input could be captured without allowing for the possibility that every aspect of the interpretation that we make of an input depends on all other aspects of the input at the same time.

(16:13):

So he wrote this paper in which he articulated this idea, and it was before he started thinking about neural networks. He was reacting to the brittleness and discreteness and difficulty of getting mutual interdependencies to be captured in explicitly discrete computational models. He tried it for several years, but two years later, I'm reading this technical report where he is saying, "This is not going to work at all." We need an approach that allows all aspects of what we're processing to influence the way we process all other aspects of it.

(16:56):

It was a kind of joint realization that a neural network would be the way to capture that In the earliest ways we thought about it, the units did correspond to things that we would cognize like particular visual features of letters or whole letters, or a whole word made out of a specific sequence of letters, or one of the two alternative meanings of a word like bat that could refer to a baseball bat or a little flying animal.

(17:28):

So by taking that perspective, you can say, well, for each of those units, its state of activation can depend on the states of activation of all the other units in this bidirectional kind of way. So if I have evidence that I'm thinking of a particular word, that word can tell me that these letters must be there. And if I have evidence of these letters, these letters can tell me that that word must be there. And if I have evidence that there's somebody standing at the plate in a baseball game and somebody uses the word bat, I have evidence that they're talking about the thing that that guy has in his hand that he's going to swing at the ball. So that's what Rummelhart's paper was about,

Nicholas Weiler (18:12):

And neural networks have been able to do really remarkable things. You go through a lot of them in the book, and I'd love to go through one or two of those with you. I'll just say that the idea that's forming in my mind as you were describing these early sort of reactions to a more computational oriented or programmatic oriented cognitive science is, very early on if you're going to teach someone computer science, a common thing to do is to ask someone to describe how to make a peanut butter and jelly sandwich in a way that a robot would be able to do.

(18:44):

And what you always find is that it is not an intuitive way for us to think. You do not usually think you have to tell the robot, look, which side of the bread to put the thing on or that you have to clean the knife in between. Or there are all these steps that if you want to spell everything out for a very stupid robot, it is very challenging, whereas a human just knows how to do it and doesn't go through all of these steps.

(19:10):

And so, what it sounds like you're saying is we learn in a different way. We think in a different way than a computer would, traditionally. There was a beautiful passage in the book that really helped crystallize this for me, and it's where you were talking about imagining how an infant would start to make sense of this chaotic world, right? Imagining a relatively blank slate. I know that's a somewhat controversial term because we have lots of evolution as well, but building up representations of, what are these flat surfaces? What is a room, what is a house, what is a bed, what is a table, and gradually building up. Well, I see these things that I start to see as trees or as animals and so on.

(19:54):

Basically what I want to ask is, can you help us understand at a high level, how a neural network model might explain how we start out not really knowing anything about the world and end up, like my children, knowing an incredible amount of detail about the difference between whale sharks and hammerhead sharks, or all the different flavors of Pokemon that they seem to know everything about. What's the process that would allow a neural network to build up to that level of understanding?

Jay McClelland (20:25):

Well, I want to say at this point that even though our current AI systems seem to exhibit a great deal of such knowledge themselves, you can ask ChatGPT to tell you about all those different kinds of sharks and so on. We actually learn far more efficiently than our machines do, and we don't know how to build the alerting rules that change the strengths of the synapses to allow our brains to exploit this information as rapidly and as effectively as we do.

Nicholas Weiler (20:57):

But I mean, I guess what I'm getting at is these sort of Rumelhart style networks could do something similar to learning these categories without getting a lot of explicit instruction. We were talking about how neural networks help you get away from a more programmatic approach where you would have to program in to a model.

(21:17):

The rule about sharks is that they have this size of fin and they live in the water, and a hammerhead has a head that has exactly this shape and so on. And you would imagine a huge number of rules you would have to learn, whereas in the neural network model that you're describing, a lot of that comes for free from just asking the network to do a particular task, like categorizing all kinds of different images, right? If the network is large enough, it starts creating distinct representations in its artificial nodes, and out of that, you get all of these features without sort of learning a set of explicit rules. Am I conceptualizing this correctly?

Jay McClelland (22:03):

I think there's a very important element of validity in what you're saying. What I was trying to get at before was that we learn two kinds of things. We learn what you were just describing, as well as what people expect us to say when we engage in discourse with them about objects.

(22:22):

I think of both of those kinds of learning as being things that emerge gradually from this process, which we can model concretely in a very explicit kind of setup. We can model concretely this process of classifying images, so actual photographs presented to our model at the level of just red, green, blue, sort of intensity values at each of tens of thousands of pixels in an array.

Nicholas Weiler (22:54):

Okay.

Jay McClelland (22:55):

That's an image. That's the input. And then on the output, we simply have units corresponding to these classes. Like, that image is an image of a cat. This image is an image of a dog. This image is an image of a gray whale. This image is an image of a dachshund.

Nicholas Weiler (23:17):

Right.

Jay McClelland (23:17):

Okay. Nobody tells the model the rules, but you do in this case have images and labels. So whenever the kid hears me say, "Oh, look, that's a dog," and it's looking at a dog, it has an image and a label, and it's going to be building up something like this neural network that we can train explicitly the way I was just describing, where all we have is images and labels.

(23:45):

But much more is going on than that too. Mommy might say this kind of dog has very fluffy, curly hair. It's called a poodle. The child might learn that dogs with curly fluffy hair are called poodles, because they learn to anticipate the words that Mommy's going to say from the words that she's previously said, and also they have the visual input of the dog at the same time.

(24:10):

The really shocking, sort of step here, is to go beyond saying the knowledge that underlies our ability to recognize the difference between different species of dogs is this implicit knowledge and can't be captured by rules, to saying that the knowledge that allows us to come out with verbal statements to each other is also not represented by any of those rules. It's also implicit in the weights and connections would've been built up over time in the discourses that we have with other people.

(24:45):

It's a similar process in the AI systems. You train image classification with images and labels. In language processing, you train them with sequences of words or things called tokens, which are often parts of words, but to just predict what the subsequent tokens are going to be, right? So nobody says there's a rule here about this. You just experienced the sequence of tokens, so that if there are rules, they're latent in the sequences. And if you've learned them in some sense, you've picked up on these dependencies, but they're all in connections in neural networks, one for classifying images and another for predicting tokens. There's a lot of integration across these streams, but neither of them necessarily depends on systems of rules.

Nicholas Weiler (25:48):

So I mean, the book dives deep into a lot of these, how would you build this model? How would it compare to what a human mind does either in infancy or as an adult, in processing the inputs that we get from the world acting on our goals and motivations and so on. And sort of suggests that you can get a lot of these things through these emergent behaviors of networks, of simplified neurons with relatively simple rules.

(26:17):

And we've talked before on the show, we talked with Deborah Gordon about how you get very intelligent behaviors from a colony of ants without any individual ant being particularly intelligent. And you use a similar example in the book, that without neurons being intelligent, you can have intelligence arise.

(26:35):

And so, the thing I wanted to ask you sort of a higher level question is how does this help you at a broader level, understand we mean when we talk about the mind? I think that people have a sense that there's something in the head. There's some person with a perspective that is evaluating our experiences and making decisions and so on. But the argument that you make is you can get almost all of the intelligent behaviors we can think of through the pattern of activation through these neural networks. So where does that leave the mind? Is that all the mind is?

Jay McClelland (27:11):

I think that it's important to begin to entertain that as an exciting possibility, while also recognizing that we do not yet fully understand how it gets to be as coherent seeming as it does.

(27:31):

I tend to think that our thoughts are a bit more structured than the way they look in today's language models, but it's only a matter of degree, and a matter of something like this mutual interaction effect that I was trying to allude to before, where things that are mutually consistent sort of flow together to create states of our mind that feel more holistic and coherent, rather than just being all of this firing of neurons. If you hear a sentence in a language that you don't know, your experience will be a sequence of sounds that leaves little residue in your memory. But it doesn't hang before the mind, and it doesn't make itself feel like you've got your head around it.

(28:30):

But if the sentence is one that makes sense to you, and expresses an idea that you already sort of have the wherewithal to understand, like a peanut butter and jelly sandwich is made by spreading peanut butter and jelly on some bread and putting the two slices of bread together, this will hang before your mind. So you could almost repeat exactly what I just said to you.

(28:56):

And this hanging before your mind depends on the fact that all the parts hang together with each other because of the previous experiences that you've had, where we express relationships and it brings to mind additional elements of the actual stickiness, maybe, of that sandwich with the stuff in between those two slices of bread. Interaction of elements constraining each other that causes this thought to hang before your mind in a coherent way, that's dependent on your prior knowledge and experience in this myriad of ways and makes it so it feels coherent and meaningful.

(29:38):

The key thing to recognize is that that only happens if you have the relevant experience. So if that sentence was in Arabic and you'd never heard Arabic before, and it was about that exact same peanut butter sandwich, you basically wouldn't have even heard it, right? So the fact that it hangs together and makes sense also causes it to hang together at the level of the sounds that you hear.

(30:04):

And this is something that was demonstrated by neuropsychologists who looked at patients who have brain damage that influences their ability to access the meanings of words. Once they've lost that ability, they can't hang on to the sequence of words in the same way. So this sort of coherence of thought is something that arises from the mutual interacting constraints between all the participating neurons that are elements of this thought at all the different levels of description we could have of it, hanging together and mutually supporting each other in ways that make our thoughts seem coherent and stable, and which I think are not fully captured in the AI systems that we currently have, and that I haven't been able to completely figure out how to build myself.

Nicholas Weiler (30:58):

I was getting this very holistic picture hanging before my mind as you were speaking, that one of the things that comes out of this neural network way of thinking about the mind is, again, it is unstructured, but it is a way of representing all the complexity of our experience, and learning patterns from that that are useful for our everyday lives. That's why we have a brain, is to learn from experience, to be able to move through the world in a way that's beneficial for us as an organism and a species.

(31:31):

And so,, what you're saying about an experience being a set of associations between whatever it is we are seeing or what we are hearing from another person and the various different sort of echoes of activation that reaches out and brings to bear, it's all of those things in a way that's somewhat structured by the structure of our brains and the different brain systems like hippocampus and amygdala and nucleus accumbens and all the things that we've talked about many times on the show. But also something that is deeply shaped by our individual experience.

(32:08):

And I had this experience many times while reading this book. It's very heady to try to think of your own mind in this way. And you brought up the ways in which the things that we can do with our minds, our experience of the world is really not captured yet, let's say, by modern AI systems as powerful as they are.

(32:27):

You listed, I think eight areas, towards the end of the book where you think our understanding of our brains could influence the next step of the evolution of these AI systems. Could you lay out one or two of those that you think are going to be really influential from what we're learning about our own minds on what we are going to do with these new AI in systems?

Jay McClelland (32:52):

Yeah. So the first one that hits me over the head very strongly, the models that we use today are trained with a hundred thousand times more language data than a human being could experience in their lifetime. You'd have to live a hundred thousand lifetimes to get the same amount of experience as a language model is getting when it learns.

Nicholas Weiler (33:18):

And a kid can speak by the time they're two.

Jay McClelland (33:20):

Yeah, I mean, they're not an expert when they're two, but by the time they're 20, they've had only 100000th of the amount of experience that like DeepSeek model that came out earlier this year had.

(33:34):

And so what I like to emphasize though, that we don't understand how our brains have succeeded in learning so much more efficiently so that we need to be doing the neuroscience to understand that, to be able to then make a contribution to fully addressing this problem in AI. So the neuroscience that remains to be done will be part of the solution here.

(34:00):

There are new breakthroughs happening in the neuroscience of learning. We had a speaker earlier this quarter who has introduced what to me is a radical new way of thinking about some of the important details of how the synaptic connections between neurons get changed. His ideas have made it possible to see how neurons could change their own contribution, an individual neuron could change its contribution, in ways that most neuroscientists had not been thinking about up to. I don't think anybody was thinking about these ideas.

Nicholas Weiler (34:37):

So maybe the individual neurons are not such simplistic units after all.

Jay McClelland (34:41):

Well, they're still simple. It turns out that it's still simple, but it's happening at a different timescale and under a different degree of depolarization, based on these latent traces that are left behind by synaptic inputs that last for much longer than people had previously thought.

(35:01):

And so, it's not like it isn't simple, but it's just happening on a different scale. That's one example though.

(35:09):

Another thing that is very important in which some of our earlier conversation already points to is that the models in use in AI today do have this input-to-output kind of characteristic, right? You present that image to the classification model, it goes forward through many, many layers. It gradually learns all the connections among those layers, but there's no reciprocal interaction backward through the network, making it so that when elements in one part of an image can influence the way you interpret the elements in the other part of the image. It's all just a feed forward process culminating in, "Oh yeah, that's a poodle."

Nicholas Weiler (35:49):

Constantly using what we're doing and seeing and thinking to predict the next thing and to influence the next thing.

Jay McClelland (35:54):

The context is, the prior and immediate context are interactively flowing in a bidirectional way, as I keep emphasizing in our brains, in ways that are not utilized in today's AI systems. So we know that those connections exist, that go in both directions in the brain.

(36:16):

There's debates among neuroscientists as to exactly what roles they play. They are largely ignored in much of neuroscience research. But I think, again, that this is a place where much, much more work is required to fully understand how the brain relies on these bidirectional interactions that Rumelhart and I were very committed to capturing in our early modeling work, that are missing. Both from the way we modeled the brain typically, and the way we use neural networks in AI. That's another important frontier in this area.

Nicholas Weiler (36:52):

Fantastic. This is such a fascinating area of thinking that we are confronted with these large language models, what feels very much like another intelligence. But we still have a lot to understand about the mechanisms of our own minds and the related mechanisms of how these systems are working, what that means about what it means to have a mind, how does the mind emerge from these simpler systems, and what it means for sort of how we think about ourselves and how we think about these AI models going forward.

(37:25):

So I really want to, again, thank you and Gourav for writing this very thought-provoking book, laying out a lot of these ideas, and thank you also for coming on the show and talking with us about it.

Jay McClelland (37:37):

Well, it's been a tremendous pleasure, Nick. And I hope to leave your audience with the idea that just as Darwin's discovery of the principle of natural selection as a basis for understanding evolution completely revolutionized the science of biology, I think the neural network perspective is beginning to emerge as one that's going to completely revolutionize the science of mind.

(38:05):

But it took a long time for Darwin's ideas to take hold, and there were very important mechanistic insights that didn't happen until much later, like the discovery of the double helix. And so, I think that the future of this kind of way of thinking is going to be as rich as the future of biology has been since Darwin and is going to be studied with many future onsights.

Nicholas Weiler (38:36):

Thanks again so much to our guest, Jay McClelland. Jay is the Lucy Stern professor in the social Sciences and a professor of psychology in the School of Humanities and Sciences at Stanford. His book is The Emergent Mind: How Intelligence Arises in People and Machines, co-authored by Gourav Suri.

(38:55):

To read more, check out the links in the show notes. If you're enjoying the show, please subscribe and share with your friends. It helps us grow as a show and bring more listeners to the frontiers of neuroscience. Also, quick thanks to listener Robby Bannon, who got in touch recently to share her thoughts on the show. We had a great conversation. And we'd love to hear from more listeners about what's working for you on the show and what you'd like to see more of. Send us an email at Neuronspodcast at Stanford.edu. From Our Neurons to Yours is produced by Michael Osborne at 14th Street Studios, with sound design by Mark Bell. I'm Nicholas Weiler. Until next time.