The Emergent Mind, page 1
To our wives, Ritika and Heidi
Contents
Preface
Part 1: Your Mind Is a Neural Network Chapter 1: An Invitation
Chapter 2: How Could the Mind Arise from the Brain?
Chapter 3: What Does a Neural Network Do?
Part 2: Activation Produces Thought and Action Chapter 4: A Neural Network as a Memory System
Chapter 5: Context Matters
Chapter 6: The Things We Do
Part 3: Knowledge and Learning: They’re in Your Connections Chapter 7: Making (and Losing) Meaning
Chapter 8: The Emergent Thinking Machine
Chapter 9: When We Learn, We Change Connections
Part 4: Extending and Applying the Neural Network Framework Chapter 10: Our Emergent Thoughts
Chapter 11: Implications of the Neural Network Framework for Ourselves and for AI
Acknowledgments
Further Reading and Chapter Notes
Illustration Credits
Index
Preface
Our brains are vast arrays of cells called neurons, animated by patterns of electrical and chemical activity that rise, fall, and rise again. Our perceptions, thoughts, decisions, and actions—the processes we will call the mind—arise from these patterns of activity.
How? How could a mind possibly arise from patterns of activity in a brain?
To us, this is one of humankind’s most enduring and most stirring questions. It concerns our essence and our place in the universe. It also concerns the possibility—and, if possible, the nature—of artificial minds. We have written The Emergent Mind to share a new kind of answer to this how question.
We are practicing scientists who have dedicated our careers to understanding the human mind. When we started our initially separate journeys, we each sought to investigate whether the operations of the mind can be understood mechanistically—the way one might seek to understand, say, an aircraft or the way a virus causes a disease. We found the existing approaches related to our questions to be vaguely specified and often too far removed from relevant facts about the underlying brain activity. We thought that grounding our explanations in the mechanistic workings of the brain could lead us to better answers to questions about ourselves.
The physicist Richard Feynman left this on his blackboard when he died: “What I cannot create, I do not understand.” His statement captures an essential aspect of our approach. We seek to build brain-like systems that re-create the phenomena of the mind that we are seeking to understand. However, re-creating a brain in all its details is impractical. Instead, we build models that abstract away from many details to help us make progress.
The models we use—called neural network models—are inspired by the vast networks of neurons in the brain. They allow us to explore how our human capacities might arise from neural activity. These models purposely ignore many of the complexities of the brain to focus on the basic processes that help us understand how our minds work. In this book, we will describe neural networks that shed light on how humans perceive, decide, form concepts, and pursue goals.
These models also help us understand answers to questions about ourselves that have puzzled us and many others. These questions often begin with words like why, what, and where. Why do we sometimes fail to act on our intentions? Why do we and others exhibit entrenched biases? What is it about us that allows us to sometimes easily see that something is true, while other times we can fail to understand? Where do our intuitions come from, and why can they often be wrong?
Remarkably, neural networks—implemented as computer programs—have become the foundation of artificial intelligence. Models we and others originally developed to understand the human mind turned out to be the basis for building artificial minds. Thus, understanding how neural networks capture our thought-like capabilities also sheds light on today’s AI systems. In this book, we will discuss the key principles of neural network models of our own minds that underpin the AI systems that are reaching and sometimes exceeding aspects of our human cognitive abilities. While our focus is on what we consider to be enduring ideas, we also selectively discuss certain rapidly evolving innovations in AI systems from the mid-2020s—particularly in the final two chapters. These innovations, while interesting and generative, may become much more developed in the coming years. Whenever possible, we have described such ideas in a way that emphasizes their underlying principles rather than their transient implementations. Our goal is to provide readers with a foundation that remains useful even as specific implementations evolve.
The book is divided into four parts. In Part 1, we begin by describing how a system can have properties that are not present in any of its parts. This phenomenon—called emergence—is central to the neural network framework of the mind, according to which the mind emerges from the interactions between simple processing units akin to brain cells. The brain cells, by themselves, cannot think, but their interactions enable a system that does think. In Part 2, we showcase how the neural network framework of the mind can help explain a diverse range of human behavior. First, we examine neural network–based accounts of memory—including its fallibility. Next, we consider our dependence on context to make sense of the world around us—including how our expectations shape our thoughts. Then, we consider our decision-making—including how our choices are sometimes rational and sometimes irrational. In Part 3, we detail how neural networks—both biological and artificial—learn from experience. We describe how learning enables our knowledge about objects and their properties and how it can support using language—especially in large language models (LLMs). Finally, in Part 4, we extend and apply the neural network view of the mind. We describe how neural networks might be useful in understanding phenomena such as formal reasoning, motivated behavior, and consciousness—aspects of the mind that have not yet been fully captured in the neural network framework. We close by discussing some implications of the neural network perspective—both for us humans and for our machines.
Throughout the book, we include interludes that are, unless noted otherwise, fanciful and fictitious conversations. For example, one conversation features Sigmund Freud talking to Adam Smith, and another features the editor of this book talking to a (fictitious) customer at a bar in New York City. We have imagined these conversations, and all the words attributed to people—whether historical or fictitious—are figments of our imaginations. We hope these interludes will enliven your consideration of the issues as much as they have enlivened our writing of this book.
We wrote this book for anyone curious about minds—both human and artificial. We do not assume any mathematical sophistication. Other than simple multiplication and addition, there is not a single equation in the book. We also do not assume any prior knowledge in the cognitive, psychological, neural, or computer sciences. At the end of the book, we provide notes with pointers to resources for those interested in learning more.
Our understanding of the mechanisms of the mind continues to evolve. There is much that remains to be discovered—but what is already known is stirring and profoundly consequential.
We invite you to embark on this journey with us. Perhaps it will enrich how you see yourself and your place in the universe we live in.
—Gaurav and Jay
Part 1
Your Mind Is a Neural Network
In Part 1, we invite you to consider the proposition that our minds are usefully conceived of as arising from interactions between brain cells that, by themselves, do not have the capabilities of the mind. We introduce the idea of emergence—a phenomenon in which the whole has properties that are not present in any of the parts—and describe how neural network models help us understand the emergence of the mind.
Chapter 1
An Invitation
When one of the authors of this book, Gaurav, was fourteen, his parents gave him the equivalent of about fifty dollars for his birthday. The cash was enough for him to buy denim bell-bottoms—absolute must-haves for any teen at the time—or enough for him to reserve a spot on a longed-for school trip with many of his friends. The problem was that he could have one, but he really, really wanted both. A decision was needed. So, that evening, he bit the bullet and chose. He would take the trip. The pants, after all, could wait. He felt sure that he was doing the right thing.
But the next morning, something unexpected happened: Gaurav woke up feeling sure he should choose the pants. Nothing related to the two options had changed, yet his decision had changed. This vacillation, repeated several times over the next few days, greatly puzzled him. At that time, he had just started working with computers, and he imagined that the mind was a sort of computer that operated based on logical principles. What kind of computer gave one answer in the evening and a different one in the morning? How could answers based on logic change for no apparent reason? And if his thoughts and preferences were not the output of logic and reason, then what kind of thing were they?
Gaurav had stumbled upon some of humankind’s enduring questions: How do our thoughts arise? Why do we do the things that we do? Can we trust what we think? More generally, what is the mind, and how does it work?
Common Conceptions of the Mind (and Their Limitations)
What is the mind? We might see it as whatever is inside of us that gives rise to our thoughts, perceptions, memories, feelings, decisions, and actions. But what is it, really? Where does it come from? Here we briefly consider several common conceptions and their limitations.
The problem is that while ascribing the mind to something eternal is a stirring and even beautiful idea, it is not really an explanation of what the mind is and how thoughts arise. Instead, it treats the mind as an ineffable entity resistant to further understanding. If the goal is to understand how the mind produces thoughts and the other things we ascribe to it, this is not an acceptable stopping point.
A second conception suggests that the mind is a set of sentence-like beliefs and desires. For example, “People with graduate degrees make more money” is an example of a belief. And correspondingly, “I want to eventually land a well-paying job” is an example of a desire. It seems reasonable that beliefs and desires can interact to produce intentions and actions. If someone asks why we decided to act in a certain way, we may refer to the beliefs and desires that appeared to drive our decision. Why did we apply to graduate school? You might answer, “To get a better-paying job.” Perhaps the operations of the mind simply involve the interactions of beliefs and desires to form goals that then guide our behavior.
One limitation with the belief–desire model of the mind is that it does not explain where these beliefs and desires come from. How do such abstract things as beliefs and desires result from physical processes taking place in our brains, and how can they give rise to physical actions, including moving, acting, and producing language? Moreover, the model does not explain why people frequently do not act according to their beliefs and desires. For example, patients often do not take medications essential to their health, and employees do not start retirement accounts crucial to their financial future. This occurs even though such individuals hold positive beliefs about the efficacy of their medications and desire the security provided by retirement accounts. And yet they fail to act.
Another conception imagines that the mind is akin to software that takes input from the outside world and applies a set of rules, perhaps provided by evolution, to those inputs. Indeed, in some cases, the operations of the mind seem to involve applications of rules: If an animal has wings and can fly, we will likely classify it as a bird; we choose a particular item on a restaurant menu because we think that it maximizes our value compared to other available items; and we predict the past tense of a recently invented verb such as fax to be faxed, in accordance with the easily stated rule that the past tense of any item classed as a verb is formed by adding ed.
A problem with the mind-as-software view is that it can’t take us very far in understanding the mind. Yes, many birds can fly, but we can still recognize a bird that cannot fly. Yes, we often choose menu items that we prefer, but our choices often feature the influence of variables that are incidental to inherent value—such as whether a menu item is toward the top of the page. And yes, we often add an ed to make past tenses, but there are many irregular verbs where this does not work (sleep, for example, becomes slept). Another crucial limitation of these models is that they did not result in artificially intelligent systems that worked. The approach we present in The Emergent Mind has proven to be far more successful.
A final conception that we frequently encounter imagines different aspects of the mind as depending on separate specialists, each residing in its own specialized region of the brain. According to this perspective, we move because of brain regions that specialize in movement, we see because of brain regions that specialize in sight, we get motivated because of brain regions that specialize in motivation, we speak because of brain regions that specialize in language—and, in some versions of such accounts, we think by relying on specialized regions that think.
It is indisputable that brain regions show some degree of specialization. The question is, what gives rise to this specialization? One approach, championed by the philosopher Jerry Fodor in his book The Modularity of Mind, is that this specialization arises due to the particular internal properties of those brain regions, selected by evolution to carry out computations specialized for the tasks they perform. While brain regions do differ from one another internally to some degree, the view we present in The Emergent Mind is that such specialization is largely a consequence of differences in connections into and out of different brain regions. For example, the part of the brain called the visual cortex plays an important role in visual perception because it receives especially strong input from the eyes. This suggests that changing the input coming into a brain region should produce a change in the function performed by that brain region. Indeed, people whose visual cortex receives no visual input because they are blind from birth repurpose this part of the brain for nonvisual tasks, such as processing auditory or tactile inputs, both of which provide some input to this brain region.
Rather than viewing the mind as a collection of rigidly specialized modules, this perspective invites us to see it as an adaptable system shaped by experience, learning, and environmental demands. This approach helps us explain the existence of what some scientists call “the visual word form area”—a brain region that seems specialized for reading printed, visually presented words. It is not plausible to suggest that evolution selected this region to specialize in reading, because reading and writing were invented about five thousand years ago. Five thousand years is far too short a time for evolution to have produced a dedicated reading module through natural selection.
Yet this brain area does specialize in reading for people who have learned how to read. Why? Reading depends heavily on distinguishing fine-grained visual details like the differences between very similar letters. The visual word form area receives strong input from other neurons in the visual parts of the brain that provide the highest sensitivity to details—and so it becomes recruited in the reading process. In individuals who do not read, this area becomes specialized for other tasks that also depend on distinguishing fine-grained details, such as recognizing different people’s faces. Such findings support our view of specialization in the brain as being shaped by experience and dependent on the inputs and outputs of brain regions rather than having dedicated functions prespecified by evolution.
We have considered several conceptions of mind, each of which has limited explanatory power. We now turn to a conception that has created a pathway to understanding our minds much more deeply and that has enabled the rise of artificial intelligence: the conception of the mind as arising in a neural network.
What Is a Neural Network?
Faced with the limitations of the common conceptions of the mind, some scientists and mathematicians sought to approach understanding the mind in a radically different way. They thought that it might be productive to begin to understand the outputs of the mind (i.e., our thoughts and actions) by tracking the signaling within the networks of neurons in our brains. The brain is composed of billions of cells called neurons. The neuron is the fundamental building block responsible for processing and transmitting information within the brain and the rest of the nervous system.
At a basic level, neurons do some very understandable things: (1) They can become activated—meaning they can generate small bursts of electricity called action potentials; (2) they can transmit signals to other neurons that they are connected with; and (3) they can build or adjust the strength of connections with other neurons.
Now, let’s try to visualize a network of neurons. It is cumbersome to depict real neurons, so we might agree to denote neurons with circles—which we might call units. Some of these units are connected to others, and we might denote these connections by drawing arrows between units. We might designate units that get activated by receiving inputs from the outside world as input units, units that send output to the outside world as output units, and units that have no contact with the outside world as hidden units. And there it is, a neural network (Figure 1.1).
