A Thousand Brains: A New Theory of Intelligence

by Jeff Hawkins

Basic Books, 2021

In his second book on neuroscience, A Thousand Brains, Jeff Hawkins builds on the ideas introduced in his first book, On Intelligence, where he proposed that the brain creates a model of the world, using it to predict the future. In A Thousand Brains, he expands on his idea with his Thousand Brains Theory, which explains how the cerebral cortex works and explores its implications for both human and machine intelligence.

Here are some of the key concepts from the book that I found particularly insightful:

Two Tenets of Neuroscience

  • Thoughts, ideas, and perceptions are the activity of neurons.

  • Everything we know is stored in the connections between neurons.

The Structure of the Neocortex: Cortical Columns

The neocortex comprises approximately 150,000 cortical columns—modular units with a similar structure and processing algorithm, regardless of location. This is likely a result of the evolution, where the cortical column as a functional unit is replicated as the neocortex expanded. By reusing the same basic structure, the brain can efficiently increase its processing capacity.

What is fascinating is that cortical columns process a wide variety of sensory inputs—visual, auditory, tactile, etc.—but the underlying processing mechanism remains the same across these modalities. The difference between the cortical columns lies not in their intrinsic function but in what they are connected to, essentially how they interface with the broader sensory and motor systems.

The Predictive BrainL Learning Through Predictions

Hawkins emphasizes that the neocortex is not a passive recipient of sensory input; it is constantly making predictions about the world. Neurons in the neocortex have thousands of synapses, and their dendrites fire in response to predicted sensory input. These predictions help the brain test its model of the world. When neurons predict correctly, they fire, while others remain inhibited. This predictive mechanism is central to how we learn and interact with our environment.

The Role of Reference Frames in Building a World Model

A key idea in Hawkins’ theory is that each cortical column has its own model of the world. These models are constructed through reference frames, which are spatial grids formed by “grid cells” and “place cells” within the column. These cells create a reference frame that allows the column to map out and track sequences of sensory input from different locations.

For example, when learning about an object like a cup, the brain doesn’t just process a single point of contact. Instead, we gather sensory input from multiple points—such as the texture of the cup, its shape, and its weight—by moving our hands or fingers around the object. This sequence of inputs is crucial for building a robust model of the object.

Distributed Knowledge Across the Neocortex

Hawkins proposes that knowledge is stored in a distributed manner across the entire neocortex. Each cortical column learns to recognize and store information about objects or concepts, but the models are not localized in a single area. Instead, different columns may have different perspectives or fragments of knowledge about the same object. These knowledge models are spread across thousands of cortical columns, creating a highly decentralized, yet cohesive, system for understanding the world.

A Singular Perception from Many Models

Despite the fact that our brain holds multiple models of the world in different cortical columns, we experience a singular, unified perception. This singular perception arises from a voting-like process, where the various models contribute their “opinions” to create a coherent understanding of what we are perceiving at any given moment.

The Interplay Between the Old Brain and the New Brain

Hawkins distinguishes between the “old brain” (which evolved earlier) and the “new brain” (the neocortex). The old brain is responsible for basic survival instincts and emotions—like fear—while the new brain handles complex cognitive tasks such as reasoning, planning, and decision-making.

While the old brain may trigger an emotional response in reaction to a perceived threat, the new brain can use logic and reasoning to mitigate or overcome these primal reactions. Hawkins suggests that emotions like fear may not be necessary for artificial intelligence unless they are specifically programmed into it. While consciousness remains an elusive concept, Hawkins compares its mystery to the pre-DNA era’s uncertainty about the nature of life.

The Nature of Intelligence

According to Hawkins, intelligence is the ability to learn and refine a model of the world. This learning process involves creating and continuously adjusting many small models of the world, each grounded in reference frames that store knowledge and generate behavior. Three key elements are needed for intelligence:

  • Embodiment: Sensory inputs that can move and change over time (e.g., physical sensors or robotic actuators).
  • An Old-Brain Equivalent: Basic, instinctive behaviors such as survival actions.
  • A Neocortex Equivalent: A system capable of general-purpose learning and adaptation.

Hawkins argues that the ability to predict and learn from sensory data is the key to intelligence in humans and machines. This dynamic model-building process allows biological and artificial systems to interact with and understand the world.

Perception as a Simulation of Reality

One of the most profound insights in the book is the idea that what we perceive is not the world itself, but rather our brain’s internal model of the world. In essence, we live in a simulation constructed by the brain. This model is constantly updated based on sensory input and predictions about the future.

Old Brain vs. New Brain: A Tug-of-War

The old brain and the new brain often work in tandem but can also come into conflict. The old brain is responsible for basic survival functions and quick emotional reactions, while the new brain handles reasoning and complex decision-making. This dynamic can sometimes result in cognitive dissonance or errors in judgment—when the old brain’s emotional responses conflict with the new brain’s logical reasoning.

Implications for Neuroscience and AI

Hawkins’ Thousand Brains Theory provides a new framework for understanding how the brain processes information, learns, and perceives the world. His theory has significant implications for both neuroscience and artificial intelligence. For example, Hawkins challenges the hierarchical model of information processing found in traditional machine learning systems, such as convolutional neural networks (CNNs), and instead suggests that the brain’s architecture is parallel and distributed across cortical columns.

Additionally, the idea that the brain learns through sequences of sensory inputs rather than static inputs could influence future AI models, particularly in areas like reinforcement learning and sequential decision-making. To more closely mimic the brain, machine learning systems may need to incorporate models that better capture dynamic and sequential patterns of input.

Conclusion

A Thousand Brains offers a fresh and compelling theory of how the brain processes information, builds models of the world, and generates intelligent behavior. By rethinking the brain as a system of distributed, parallel processing units—cortical columns—Hawkins provides a new perspective on understanding human cognition and the future of artificial intelligence. His theory opens up new possibilities for building intelligent machines that learn, adapt, and interact with the world in ways that more closely resemble human intelligence.