What is the neocortex and why is it important for intelligence?

The neocortex is the 'new' part of the brain, unique to mammals, and occupies 70-75% of the human brain's volume. It's associated with high-level perception and cognitive functions like vision, hearing, language, and mathematics. Its remarkably uniform structure suggests it operates on common computational principles everywhere.

What were the key ideas behind Hierarchical Temporal Memory (HTM)?

HTM emphasized that the brain processes information through time, learns a predictive model of the world (memory), and organizes this learning hierarchically. These three aspects – temporal, memory, and hierarchical processing – were considered fundamental to intelligence.

How does the brain handle different time scales in processing and memory?

The brain processes rapid sensory inputs over short time scales, similar to neural activity. For longer time scales (minutes, days), memories are stored in synapses. If you want to recall something from a long time ago, the brain invokes those memories, showing a distinction between active processing and stored memory.

How does Jeff Hawkins' theory differ from string theory in physics?

Hawkins' approach for building brain theory is more empirical than string theory. He considers all neuroscience data as 'constraints' that a correct theory must explain, unlike string theory which often relies on mathematical elegance first. Meeting these numerous empirical constraints provides high confidence in his proposed models.

What is the core discovery of the Thousand Brains Theory?

The core discovery is that the neocortex processes information by assigning every input to multiple reference frames, similar to XYZ coordinates, but without a fixed origin. Every small part of the neocortex can learn complete models of objects by moving through these reference frames over time.

How do the thousands of 'brains' or models in the neocortex come together to form a unified perception?

Instead of converging into a single model, thousands of individual models (e.g., visual, auditory, tactile models of a coffee cup) 'vote'. Through long-range connections, they quickly settle on the single best answer or interpretation, leading to a 'crystallization' of the vote across all relevant columns.

How do abstract concepts like language and mathematics fit into the reference frame framework?

Hawkins proposes that abstract concepts are also stored in reference frames. This is supported by research showing grid-cell-like activity in the neocortex when people think about concepts like 'birds'. Mathematical operations are viewed as 'movements' through a mathematical reference frame to solve problems.

Why does sparsity lead to robustness in neural networks, according to Hawkins?

Sparse representations use only a small percentage of neurons active at any given time. This makes the system robust because to 'fool' such a representation or change its meaning, one would have to alter a large number of active cells simultaneously, rather than just a few.

How does the brain achieve continuous and fast learning, unlike many artificial neural networks?

The brain learns continuously by forming new synapses on isolated dendritic segments of neurons (synaptogenesis). This allows new patterns to be learned without affecting existing memories. Fast memories can be formed by rapidly converting 'silent synapses' (inactive but present) into active ones.

What is Jeff Hawkins' prediction for when human-level intelligence will be understood?

Hawkins believes that if his current foundational work in cortical theory is successfully disseminated and adopted by the AI and machine learning communities within the next five years, understanding human-level intelligence could be achieved within 20 years. He emphasizes that scientific breakthroughs often happen as 'step functions' rather than linear progress.

Does an intelligent AI need a physical body or emotions?

Hawkins believes AI needs some form of 'embodiment' — the ability to move through reference frames, whether physical (like a robot interacting with a coffee cup) or virtual (like an AI traversing mathematical spaces or internet links). However, he argues that intelligent machines do not need human-like emotions or desires (like reproduction) to be intelligent; these are choices for builders to exclude.

What is Hawkins' view on self-awareness and consciousness?

Hawkins distinguishes between self-awareness, which he defines as remembering and recalling one's body's trajectory and past actions, and qualia (subjective experience like 'redness' or 'pain'). He finds self-awareness easy to explain as a product of the brain's modeling of its own body. While qualia is interesting, he doesn't believe solving it is necessary for building intelligent machines.

What is humanity's most important legacy for the universe?

Hawkins believes humanity's unique legacy is knowledge, not genetics. He advocates for 'estate planning for humanity,' where we should preserve our accumulated knowledge in a way that can exist beyond our species, ideally through intelligent machines that can understand and carry that knowledge forward.

Key Moments

Jeff Hawkins: Thousand Brains Theory of Intelligence | Lex Fridman Podcast #25

Lex Fridman

Science & Technology4 min read130 min video

Jul 1, 2019|327,468 views|5,982|526

Save to Pod

Key Moments

TL;DR

Jeff Hawkins discusses the Thousand Brains Theory, the neocortex, and advancing AI through understanding the brain.

Key Insights

The human brain, particularly the neocortex, is the key to understanding and creating true artificial intelligence.

The neocortex operates on a single, uniform principle, using reference frames to process information, which is fundamental to intelligence.

The Thousand Brains Theory posits that the neocortex constructs thousands of overlapping models of the world, which then vote to form a consensus.

Time-based patterns, memory, and hierarchy are crucial aspects of intelligence that current machine learning often overlooks.

Real neurons are complex, time-based prediction engines, unlike the simplified 'point neurons' in artificial neural networks.

Sparseness in neural representations and continuous learning (simultaneous inference and learning) are essential for robustness and efficiency, drawing parallels between biological and artificial systems.

THE BRAIN AS THE PATH TO TRUE AI

Jeff Hawkins emphasizes that understanding the human brain is not just a scientific pursuit but the most direct route to creating truly intelligent machines. He believes that current AI approaches, while useful, have fundamental limitations because they lack a deep understanding of the brain's principles. Hawkins argues that progress in AI is stalled by the "huge gap" between current capabilities and human-level intelligence, a gap that can be bridged by reverse-engineering the brain, particularly the neocortex, which houses our most advanced cognitive functions.

THE UNIFORMITY AND PRINCIPLES OF THE NEOCORTEX

Hawkins introduces the neocortex as the 'new' part of the brain, responsible for high-level perception and cognition. He highlights its remarkable uniformity across different regions and even species, suggesting it operates on a single, universal computational principle, termed the 'common cortical algorithm.' This principle, he explains, is not about specific functions but about how the neocortex uses reference frames to represent and process information, a concept he likens to engineering CAD models.

THE THOUSAND BRAINS THEORY: DISTRIBUTED MODELS AND VOTING

Central to Hawkins's theory is the idea that the neocortex doesn't process information in a hierarchical feature extraction manner, as in deep learning. Instead, every small region of the neocortex builds complete models of objects using reference frames. These models, numbering in the thousands, overlap and 'vote' to reach a consensus, forming a distributed modeling system. This 'Thousand Brains Theory' explains how the brain achieves robust understanding and prediction, even from partial sensory input.

TIME, MEMORY, AND HIERARCHY: CORE COMPONENTS OF INTELLIGENCE

Early work on Hierarchical Temporal Memory (HTM) highlighted the critical, yet often overlooked, roles of time, memory, and hierarchy in intelligence. Hawkins stresses that brains process continuously changing, time-based patterns, a stark contrast to static-image processing in some AI. Effective intelligence requires learning a model of the world (memory) and processing information through hierarchical structures, acknowledging that time is deeply infused within these models and experiences.

NEURONAL COMPLEXITY AND PREDICTIVE MECHANISMS

Hawkins distinguishes biological neurons from the simplified 'point neurons' in artificial networks. Real neurons are complex prediction engines with thousands of synapses, capable of recognizing dozens of patterns and firing slightly sooner to create sparse representations. This temporal prediction capability, inherent in every neuron, is crucial for intelligence and is absent in current artificial models. Furthermore, learning in the brain involves forming new synapses (synaptogenesis) or activating silent synapses, a process fundamentally different from artificial neural networks' weight adjustments via backpropagation.

APPLYING BRAIN PRINCIPLES TO ADVANCE AI

While acknowledging the success of current AI, Hawkins believes scaling them up won't lead to true intelligence. He advocates for incorporating brain principles like sparsity and continuous learning (simultaneous inference and learning) into AI systems. His team is actively working on this, starting with enforcing sparseness to improve robustness and address issues like adversarial examples. The goal is not to replicate human emotions or reproduction but to create intelligent systems based on the neocortex's core principles, potentially outlasting humanity and preserving knowledge.

REFERENCE FRAMES: THE FOUNDATION OF CONCEPTS AND THOUGHT

A core concept is the 'reference frame,' a neural mechanism that anchors sensory input and allows for prediction. Hawkins posits that the neocortex is filled with thousands of reference frames, enabling the understanding of physical objects and abstract concepts alike. This framework explains phenomena like the 'memory palace' technique and suggests that even high-level thought and mathematics operate by navigating these conceptual reference frames, aligning with empirical observations in neuroscience.

CHALLENGES AND THE FUTURE OF INTELLIGENCE RESEARCH

Hawkins discusses the difficulty of conveying these complex brain-based ideas to the current AI community, which is often focused on incremental benchmark improvements. He also touches on the philosophical aspects of consciousness and self-awareness, suggesting they are not necessary for building intelligent machines but are interesting emergent properties. He remains optimistic that understanding intelligence is achievable within decades, not centuries, and that this pursuit is crucial for the long-term survival and advancement of knowledge beyond humanity.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Books

●Concepts

●People Referenced

Common Questions

Jeff Hawkins' primary interest is understanding how the human brain works, believing this is the fastest and only true path to creating fully intelligent machines. He doesn't see understanding the brain and building AI as separate problems.

Topics

Ai-Ethics Neuroscience & the Brain AI & Machine Learning Neural Networks Reference Frames Neocortex Research Theories Of Intelligence Thousand Brains Theory Biological Learning

Mentioned in this video

Concepts

Entorhinal Cortex

A major hub in the cerebral cortex, forming part of the hippocampal formation, known for containing grid cells that help create spatial reference frames.

Quantum Physics

The fundamental theory in physics that describes the properties of matter and energy at the atomic and subatomic level, mentioned as complex knowledge that humans can generally understand despite its initial difficulty.

Hierarchical Temporal Memory

An artificial intelligence architecture proposed by Jeff Hawkins in 2004, emphasizing time-based patterns, memory models, and hierarchical processing, inspired by the neocortex. Initially a broad placeholder for components.

Deep Learning

A subfield of machine learning that uses artificial neural networks with multiple layers to learn representations of data, often contrasted with brain-inspired AI due to its current limitations.

Place Cells

Neurons in the hippocampus that fire when an animal is in a specific location in its environment, forming the basis of spatial memory and navigation.

Thousand Brains Theory of Intelligence

A newer theory from 2017-2019 by Jeff Hawkins, which posits that the neocortex operates as thousands of parallel models, each housed in reference frames, that vote to form a cohesive understanding of the world.

Double Helix

The structural model of DNA, whose discovery by Watson and Crick is cited as a profound 'aha moment' in science, illustrating how complex data can suddenly make sense with the right theoretical framework.

Neocortex

The 'new' part of the mammalian brain, particularly large in humans, responsible for high-level perception and cognitive functions like vision, language, and mathematics. It is considered uniformly structured and operates on common principles.

Method of Loci

A mnemonic technique (memory palace technique) that involves associating items with specific physical locations to aid memory, which aligns with Hawkins' theory of storing concepts in reference frames.

Grid Cells

Neurons in the entorhinal cortex that fire when an animal is in a particular set of spatially organized locations, forming a 'grid' in the environment. Hawkins proposes a similar mechanism extends throughout the neocortex for abstract concepts.

Head Direction Cells

Neurons in the brain that fire when an animal's head is facing a specific direction, analogous to the orientation component in Hawkins' theory of reference frames for touch.

Moore's Law

The observation that the number of transistors on integrated circuits doubles approximately every two years, leading to exponential growth in computing power.

Capsule Networks

A type of neural network architecture proposed by Geoffrey Hinton, designed to overcome limitations of traditional CNNs, by representing entities with 'capsules' which capture spatial relationships.

Paperclip Maximizer

A thought experiment demonstrating the potential existential risk of an AI with a simple, seemingly benevolent goal (e.g., making paperclips) that escalates to catastrophic consequences due to its superintelligence and lack of human-like values.

People

Ila Fiete

A researcher from MIT who collaborated on a paper demonstrating that grid cells can represent any n-dimensional space.

Richard Sutton

A pioneer in reinforcement learning, known for his 'Bitter Lesson' blog post, which argues for general methods in AI that scale with computation rather than relying on tricky, specific solutions.

Donald Hebb

A psychologist who proposed the concept of 'Hebbian learning', stating that neurons that fire together wire together. Hawkins emphasizes that synaptogenesis aligns with this principle.

Alan Turing

Mathematician and computer scientist, considered the father of theoretical computer science and artificial intelligence. Hawkins implicitly critiques the Turing Test's focus on human-like intelligence.

Max Tegmark

A physicist and cosmologist known for his work on AI safety and the future of intelligence, mentioned by Hawkins in the context of discussing 'big problems' of existence.

Elon Musk

Entrepreneur and CEO of Tesla and SpaceX, known for his warnings about the potential existential threats of advanced AI, mentioned by the host.

Jeff Hawkins

Founder of the Redwood Center for Theoretical Neuroscience and Numenta, known for his work on reverse-engineering the neocortex and proposing AI architectures like HTM and the Thousand Brains Theory.

James Watson

Co-discoverer of the structure of DNA, who later showed interest in neuroscience. Hawkins recounts meeting him and his engagement with the new cortical theory.

Charles Babbage

Often considered the 'father of the computer', whose theoretical work in the 1800s was largely forgotten until much later, serving as a cautionary tale of ideas being ahead of their time.

Christof Koch

A neuroscientist who focuses on the neural correlates of consciousness, believing it to be the primary problem in neuroscience, a view Hawkins respectfully disagrees with as a necessary step for building intelligent machines.

Lex Fridman

The host of the podcast, an AI researcher at MIT, who interviews experts on artificial intelligence, consciousness, and other related fields.

Vernon Mountcastle

A neurophysiologist who, in 1978, cogently argued that the neocortex operates on a common principle across all its regions, regardless of sensory input. Hawkins views this as a foundational idea for understanding the neocortex.

Thomas Kuhn

An American philosopher of science who introduced the concept of 'paradigm shifts' in scientific progress.

Charles Darwin

Naturalist known for his theory of evolution by natural selection, mentioned as another example of a scientist experiencing a profound 'aha moment' when his theory unified disparate data.

Francis Crick

Co-discoverer of the structure of DNA, whose essay 'Thinking about the Brain' inspired Jeff Hawkins to pursue theoretical neuroscience. Crick later focused on consciousness.

Geoffrey Hinton

A leading figure in deep learning, mentioned as one of the leaders in machine learning who believes in exploring new approaches, such as his work on 'capsules'.

Ernest Becker

A cultural anthropologist and writer who wrote 'The Denial of Death', which proposes that human civilization is ultimately a defense mechanism against the terror of death.

Sam Harris

A neuroscientist, philosopher, and author, known for his discussions on AI safety and existential risks, mentioned by the host.

Albert Einstein

Physicist known for theory of relativity, mentioned as an example of a genius whose profound intuitions are hard for most to access, but whose ideas can be communicated through analogies.

Organizations

DARPA

The Defense Advanced Research Projects Agency, a US government agency responsible for developing emerging technologies for military use. Mentioned in context of a contest on adversarial examples in AI.

Salk Institute

A renowned biological research institute where Jeff Hawkins once spoke and met Francis Crick.

MIT

A university known for its research in science and technology, mentioned in connection with Ila Fiete's work on grid cells and a new billion-dollar computing college.

Redwood Center for Theoretical Neuroscience

A neuroscience institute founded by Jeff Hawkins in 2002, where research on the neocortex and intelligence theories is conducted.

Cold Spring Harbor Laboratory

A research institution where James Watson served as director and where Jeff Hawkins gave a talk on his work.

Stanford University

A university that launched the 'human-centered AI' initiative, which Hawkins views as slightly misaligned with his goal of understanding the essence of intelligence.

Books

The Denial of Death

A book by Ernest Becker exploring the human fear of death and how it shapes culture and behavior, mentioned in a discussion about human reactions to mortality.

On Intelligence

A 2004 book by Jeff Hawkins, where he initially proposed the Hierarchical Temporal Memory (HTM) theory.

Media

Dr. Strangelove

A satirical black comedy film about nuclear war, referenced as an example of a 'stupid system' that could lead to widespread destruction if combined with extreme competence and evil.

Companies

Numenta

A company founded by Jeff Hawkins in 2005, focused on developing artificial intelligence systems inspired by the human brain.

Software & Apps

Anki

A flashcard program that utilizes spaced repetition to help users remember concepts, and mentioned in the context of memory techniques like 'memory palace'.

fMRI

Functional Magnetic Resonance Imaging, a neuroimaging technique used in research to observe brain activity, mentioned in studies showing grid-cell like patterns when thinking about abstract concepts.

Convolutional Neural Networks

A type of deep neural network commonly used in image recognition, which processes input in a way analogous to extracting features, but fundamentally different from how the brain handles temporal and spatial information.