How does the brain learn and make predictions according to Jeff Hawkins?

Hawkins proposes that most predictions occur within individual neurons, utilizing thousands of synapses. Neurons enter a 'predictive state,' priming them to fire slightly sooner if their prediction aligns with incoming data. When predictions are incorrect, a larger ensemble of neurons activates, drawing our attention to the discrepancy.

What is the theory of cortical columns?

The theory, pioneered by Vernon Mountcastle, suggests that the neocortex is organized into roughly 150,000 processing units called cortical columns. Each column performs a common algorithm, regardless of whether it processes visual, auditory, or other sensory information, implying a unified approach to intelligence within the cortex.

Why are reference frames crucial for intelligence?

Reference frames are essential because they provide a means to note location, which is fundamental for learning and understanding. To build models of the world, such as recognizing a coffee cup by touch, the brain needs to know the relative position of sensors (like a finger) to objects, which is what reference frames enable.

What principles does Jeff Hawkins believe are necessary for true Artificial General Intelligence (AGI)?

Hawkins outlines several key principles: embodiment (the ability to move sensors in the world, physically or virtually), the use of reference frames for organizing information, and a distributed intelligence system, similar to the 150,000 'modeling engines' in the human neocortex.

What is Jeff Hawkins's perspective on AI risk and the alignment problem?

Hawkins is optimistic, believing that the intelligence in the neocortex is primarily a benign modeling system, separate from the brain's motivational drives. He suggests that future AI can be built without replicating these drives, thus avoiding inherent motivations to act against human interests or the 'alignment problem'.

How does Jeff Hawkins differentiate the neocortex from the rest of the brain regarding motivation?

Hawkins likens the neocortex to a map – a powerful tool that understands the world but lacks inherent motivation. He posits that the rest of the brain provides the drives and motivations, and AI can be built to possess intelligence without necessarily inheriting these drives, thereby mitigating alignment risks.

Key Moments

The Future of Intelligence: A Conversation with Jeff Hawkins (Episode #255)

Sam Harris

Science & Technology4 min read59 min video

Jul 9, 2021|67,164 views|1,161|389

sam harris sam harris podcast waking up podcast waking up waking up with sam harris sam harris waking up waking up sam harris sam harris jordan peterson sam harris joe rogan author neuroscientist philosopher

Save to Pod

Key Moments

TL;DR

Jeff Hawkins and Sam Harris discuss intelligence, brain function, and AI risks.

Key Insights

Intelligence is defined as the neocortex's ability to build and use a model of the world for planning and action.

The brain constantly makes predictions, and these predictions are often processed within individual neurons, not accessible to consciousness.

The neocortex is organized into "cortical columns," each acting as a modeling engine, and the brain functions as a distributed system of these columns voting together.

Thought can be understood as movement within conceptual reference frames, analogous to physical movement in the real world.

Emerging AI systems are powerful pattern classifiers but lack true general intelligence, which requires principles derived from brain function.

Jeff Hawkins is optimistic about AI risk, believing that intelligent machines can be built without the inherent motivations and drives found in biological brains, thus mitigating the alignment problem.

A Unique Path to Neuroscience

Jeff Hawkins, with a background in electrical engineering and entrepreneurial success in handheld computing (Palm, Handspring), shares his unconventional journey into neuroscience. Driven by a deep curiosity about how the brain works, he founded Numenta and the Redwood Neuroscience Institute to pursue large-scale theories of cortical function, often self-funding due to difficulties in securing traditional research grants for ambitious theoretical work.

Defining Intelligence as World Modeling

Hawkins defines intelligence primarily through the function of the neocortex, which constitutes about 70% of the human brain. He posits that a core aspect of intelligence is the ability to learn and maintain an internal model of the world. This model enables recognition, action, and planning by allowing the brain to predict future states based on current input and internal representations.

The Predictive Nature of Neural Processing

A key insight is that the brain is continuously making predictions about sensory input, often unconsciously. These predictions are fundamental to how neurons function, with a significant portion of their synapses dedicated to preparing for expected patterns. When predictions are accurate, only a subset of neurons activate; however, prediction errors trigger broader neuronal activation, drawing attention to anomalies.

Cortical Columns and the Algorithm of the Neocortex

Drawing on Vernon Mountcastle's work, Hawkins explains the neocortex's organization into nearly identical 'cortical columns.' Each column is proposed to be a fundamental processing unit, executing a common algorithm. Despite their microscopic complexity, these columns are viewed as the core components for learning sensory-motor models and are thought to operate ubiquitously across different cortical areas, suggesting a unified principle of operation.

Reference Frames and Thought as Conceptual Movement

Reference frames are presented as crucial for understanding spatial relationships and organizing knowledge. Hawkins argues that every cortical column uses reference frames to build models of its input. Thought itself is described as movement through conceptual reference frames, an abstract form of movement that allows us to navigate and recall stored information, whether it's related to physical objects or abstract concepts like mathematics or politics.

The "Thousand Brains" Theory and Distributed Intelligence

The title of Hawkins' book, "A Thousand Brains," reflects the idea that the neocortex contains approximately 150,000 cortical columns, each acting as an independent modeling system. The brain's overall intelligence emerges from the collective 'voting' of these distributed systems. This distributed nature allows for robust sensory integration and complex processing, challenging the notion of a single, centralized intelligent agent within the brain.

Building Artificial General Intelligence

To achieve true Artificial General Intelligence (AGI), Hawkins suggests focusing on replicating the principles of the neocortex. Key requirements include embodiment (or a form of 'movement' in a representational space), the use of reference frames for organizing knowledge, and a distributed architecture. He believes that understanding how brains model the world provides the essential blueprint for creating intelligent machines.

Addressing AI Risk and the Alignment Problem

Hawkins expresses optimism regarding AI risk, particularly the 'alignment problem.' He differentiates the neocortex's world-modeling function from the motivational drives of older brain structures. He argues that intelligent machines can be built to possess intelligence without inherent desires or motivations, functioning more like a tool (e.g., a map) that can be used for various purposes, thereby avoiding the existential risks often associated with misaligned AI goals.

The Nature of Belief and Falsehood

The conversation touches upon cognitive biases like the "illusory truth effect," where the mere repeated exposure to a statement, even a false one, can increase belief. This phenomenon, particularly prevalent in language-based knowledge acquisition, highlights the challenge of distinguishing truth from falsehood and how our models of the world can become distorted. This also relates to how the brain processes propositions, often by attempting to model them as true before verification or falsification.

Challenges in Understanding Cognition

Sam Harris raises concerns about the sharp distinction Hawkins draws between reason and emotion, suggesting that these aspects may be more intertwined than believed. The complexity of human cognition, including the default acceptance of propositions and the subtle influence of biases, indicates that building truly aligned AI may be more challenging than simply mimicking neocortical structures, as the boundary between neutral modeling and motivated cognition can be blurry.

Mentioned in This Episode

●Companies

●Organizations

●Books

●Concepts

●People Referenced

Building Intelligent Machines: Key Principles from Neuroscience

Practical takeaways from this episode

Do This

Replicate the principles of the neocortex's modeling system.

Ensure AI systems have embodiment or a way to move sensors in the world (physical or virtual).

Organize information using reference frames.

Design AI as a distributed intelligence system, not a single monolithic entity.

Focus on the neocortex's function, not necessarily replicating the entire brain's motivational systems.

Avoid This

Assume current AI systems are truly intelligent; they are primarily pattern classifiers.

Replicate the 'old parts' of the brain (like brainstem or limbic system) in AI if not necessary for intelligence itself.

Underestimate the importance of movement and embodiment (even virtual) for learning.

Overlook the potential for language to introduce false beliefs into AI models.

Succumb to the idea that intelligence is inherently tied to human emotions or motivations.

Common Questions

Jeff Hawkins defines intelligence as the ability to learn and use an internal model of the world. This model allows us to recognize our surroundings, act effectively, and plan for the future, essentially understanding the world through these learned representations.

Topics

Neuroscience & the Brain Technology & Innovation AI & Machine Learning Intelligence Modeling Cortical Columns Reference Frames Artificial General Intelligence AI Alignment

Mentioned in this video

Books

A Thousand Brains: A New Theory of Intelligence

Jeff Hawkins's most recent book discussing neuroscience and AI.

On Intelligence

Jeff Hawkins's first book on intelligence.

Concepts

Illusory Truth Effect

The phenomenon where repeating information, even false information, can increase belief in it. This is discussed in the context of how language can lead to false beliefs.

Cortical Column

The proposed fundamental processing unit of the neocortex, characterized by a common algorithm operating across different sensory and motor functions.

Reference Frame

A way for the brain to note location, crucial for building models of the world, particularly in sensory-motor tasks. It's proposed to exist in every cortical column and be fundamental to all knowledge storage and retrieval.

Turing machine

Mentioned as an analogy for the fundamental principles of computation, similar to how intelligence is a system working on principles rather than specific tasks.

People

Francis Crick

A scientist whose article on brains inspired Jeff Hawkins to pursue neuroscience.

Nick Bostrom

A philosopher whose work on AI risk informs Sam Harris's perspective.

Stuart Russell

A researcher in AI whose views on AI risk inform Sam Harris's perspective.

Sam Harris

Host of the Making Sense Podcast and the interviewer in this conversation.

Béla Julesz

Mentioned by Sam Harris as a scientist who worked on visual perception and stereo vision, though the transcript garbles the name to 'Bákurita'. Hawkins does not recall the name but discusses a similar experiment.

Eliezer Yudkowsky

A researcher known for his work on AI alignment and existential risk, whose views inform Sam Harris's perspective.

Vernon Mountcastle

A scientist who proposed the idea of a common algorithm in the neocortex and the concept of cortical columns.

Antonio Damasio

A neuroscientist whose work on emotion and reason is referenced in the conversation.

Lisa Feldman Barrett

A neuroscientist known for her work on emotion, referenced by Sam Harris.

Media

Making Sense Podcast

The podcast hosted by Sam Harris where this conversation takes place. The episode number is #255.

Organizations