What is the Blue Brain Project and its goal?

The Blue Brain Project attempted to simulate a cat brain by incorporating detailed biophysical models of neurons and their statistical connections observed from experiments. The goal was to understand brain function by simulating it directly, though George expresses skepticism about this approach without a guiding functional theory.

How do detailed neuron models fit into brain understanding?

While sophisticated models of individual neurons (like Hodgkin-Huxley) can accurately replicate their spiking dynamics and biophysical mechanisms, understanding a single neuron's function does not automatically explain how millions of them form a functioning system, similar to how knowing a transistor doesn't explain a microprocessor.

How far along are we in understanding the brain's intelligence?

Predicting timelines for understanding the brain is difficult, akin to early predictions about flight. George believes progress comes from iteratively building AI models based on neuroscience insights, using functional models to fill in missing biological pieces, and informing new experiments.

What role do feedback connections play in visual processing?

The brain operates by constantly building a model of the world and projecting expectations onto sensory input. Feedback connections are crucial for this process, allowing the brain to iteratively infer the best explanation for a stimulus, rather than just a passive feed-forward interpretation.

What are cortical microcircuits and how do they encode concepts?

Cortical microcircuits refer to the intricate connections within and between layers of a cortical column. Dileep George hypothesizes that each cortical column encodes a concept (e.g., presence of an edge or object), and connections between columns encode relationships between these concepts. Neurons within the columns and the thalamus then perform inference computations.

What are Recursive Cortical Networks (RCNs) and their key features?

RCNs are brain-inspired vision models that go beyond feed-forward processing. They incorporate feedback and lateral connections, are designed for top-down controllability, allow for dynamic inference, and factorize images into separate entities like foreground/background and shape/texture, derived from natural signal properties.

Why did Dileep George's team use captchas to test RCN?

Captchas are by design easy for humans but hard for computers, requiring strong out-of-distribution generalization and dynamic inference to resolve ambiguities (like crowded characters). They serve as an excellent benchmark for understanding human-like perception with minimal training data.

How do human-like concepts and abstract reasoning relate to language?

Dileep George argues that grounded, pre-verbal conceptual understanding, learned through interaction with the world via visual and motor systems, is a prerequisite for true language understanding. Language then acts as a 'simulation control' mechanism to access and manipulate these internal world models.

What are the limitations of large language models like GPT-3?

While impressive for text generation and coherence, GPT-3 fundamentally lacks a 'world model' and cannot run simulations in its head to verify real-world truths or perform counterfactual reasoning. Scaling up transformer-based architectures for text modeling alone is unlikely to achieve true AGI or natural language understanding.

What are the current ideas on how episodic memory is formed in the brain?

Episodic memory involves storing unique, single-occurrence events separate from statistical patterns. The hippocampus is hypothesized to sequence pointers to these experiences, which can then be indexed by the cortex's statistical model of the world. Replaying these memories through the cortex helps evaluate and apply past experiences.

What are the promises and challenges of Brain-Computer Interfaces (BCIs) and Neuralink?

BCIs, including endeavors like Neuralink, hold immense promise for controlling external limbs, prosthetics, and even writing information into the brain, offering therapeutic benefits. However, significant challenges remain in safe, long-term surgical implantation, understanding brain adaptation to interfaces, and the potential for unintended neurological effects.

What's the relationship between consciousness, self-awareness, and building intelligent systems?

While not a daily guiding principle for engineering AI, self-modeling (applying the same world-modeling apparatus to oneself) is suggested as a pathway to consciousness. The human experience of death and mortality creates urgency and shapes goal generation, which might be a fundamental property of human-level intelligence, but not necessarily for copyable AI systems.

What career path is best for someone wanting to build a brain-inspired AI?

For those aiming to build intelligent systems inspired by the brain, a path through computer science, electrical engineering, machine learning, and AI is recommended. Self-study in neuroscience is crucial. Electrical engineering is particularly strong due to its foundations in circuits, information theory, and signal processing.

Key Moments

Dileep George: Brain-Inspired AI | Lex Fridman Podcast #115

Lex Fridman

Science & Technology5 min read131 min video

Aug 14, 2020|126,045 views|2,410|221

dileep george artificial intelligence agi ai ai podcast artificial intelligence podcast lex fridman lex podcast lex mit lex ai lex jre mit ai

Save to Pod

Key Moments

TL;DR

Dileep George discusses brain-inspired AI, emphasizing understanding brain principles for AI, visual perception, and cognitive architectures.

Key Insights

Building AI requires understanding fundamental brain principles, not just simulating neurons.

Human vision involves significant feedback loops, with perception being a model of the world projected onto sensory input.

Recursive Cortical Networks (RCNs) offer a brain-inspired approach to computer vision, focusing on joint understanding of perception and cognition with top-down controllability.

Captchas are valuable benchmarks for studying robust generalization and human-like perception, testing systems that learn with minimal data from novel distributions.

The brain's 'natural signals' properties inform AI architectures; conversely, engineering tricks like convolutional invariance are deviations for efficiency.

A promising path for AI involves building structured models with world representations, enabling dynamic inference and grounding language in simulated experiences.

THE NECESSITY OF BRAIN PRINCIPLES FOR AI

Dileep George argues that engineering intelligence requires understanding core brain principles, rather than solely simulating biological components. He highlights that while brain-inspired AI is often over-hyped, the fundamental insights gained from neuroscience are invaluable. The brain itself serves as an existence proof for complex intelligence. Projects like the Blue Brain Project, which attempt to build a brain by simulating detailed neuronal models without a guiding theory of function, are critiqued for lacking a computational framework to debug or understand emergent behaviors. This mirrors an analogy of trying to build a microprocessor by perfectly modeling a single transistor without understanding logic gates.

THE ROLE OF FEEDBACK AND WORLD MODELS IN VISION

George elaborates on the visual cortex, emphasizing that it's significantly influenced by feedback connections, contrary to a purely feed-forward deep learning model. Our visual system constructs a model of the world and constantly projects expectations onto incoming sensory data. Perception is thus a synthesis of this internal model and external input. This model-based inference explains phenomena like visual illusions, where the brain 'hallucinates' missing information, such as edges in the Kanizsa triangle, based on learned patterns and context. This iterative process of prediction and error correction is central to how we interpret the world.

RECURSIVE CORTICAL NETWORKS FOR COMPUTER VISION

The Recursive Cortical Network (RCN) model, developed by George's team, exemplifies a brain-inspired approach to computer vision. It treats perception and cognition as interconnected, not separate, emphasizing top-down controllability. This means the system can manipulate its internal visual knowledge, simulating scenarios and exploring details imaginatively, though not necessarily photorealistically. The RCN architecture factors out distinct elements like objects from backgrounds, and shapes from textures, deriving these priors from the statistical properties of natural signals – signals the brain is finely tuned to process, unlike artificial signals like QR codes.

INFERENCE, CAPTCHAS, AND HUMAN-LIKE LEARNING

A key differentiator of RCNs is their ability to perform dynamic inference, integrating local evidence with global context to form the best explanation. This contrasts with standard neural networks which often rely on 'amortized inference' from massive training data. Captchas, designed to be easy for humans but hard for computers, serve as a critical benchmark. RCNs demonstrate strong generalization and robustness, solving captchas with minimal training data, mimicking human ability to adapt to novel variations. This capability highlights limitations in current deep learning systems regarding out-of-distribution generalization and common-sense reasoning, essential for truly understanding concepts like the letter 'A' in all its forms.

STRUCTURAL ASSUMPTIONS AND COMPUTATIONAL TRICKS

George distinguishes between fundamental brain principles and engineering tricks. While Convolutional Neural Networks (CNNs) utilize translational invariance for efficiency, a trick useful for static images, the brain's visual system, with its fovea and differential processing of peripheral vision, employs different strategies. Understanding these underlying principles allows AI researchers to deviate from strict biological mimicry when beneficial, like using CNNs for specific tasks while acknowledging their limitations. The core idea is to understand 'how to deviate' effectively, rather than slavishly replicate biological mechanisms. This pursuit of understanding drives the development of more advanced AI models.

COGNITION, LANGUAGE, AND WORLD MODELS

Beyond perception, George's work explores cognitive architectures that learn concepts and reason. He proposes that language is a control mechanism for an internal simulation of the world built by perceptual and motor systems. This perspective challenges Chomsky's view that language is primary, suggesting instead that grounded, pre-verbal concepts derived from sensory-motor interaction are foundational. Future AI systems, he believes, will connect language to these grounded simulations, enabling true understanding rather than just sophisticated text pattern matching like large language models (LLMs) such as GPT-3, which, despite their impressive scale, lack robust world models and the ability to perform true counterfactual reasoning.

THE FUTURE OF AI: STRUCTURE, DYNAMIC INFERENCE, AND MEMORY

George advocates for AI models with more inherent structure, enabling dynamic inference rather than relying solely on brute-force training. He believes future AI will integrate properties of neural networks and graphical models, like Graph Neural Networks. Key to advancing towards Artificial General Intelligence (AGI) is the development of world models and sophisticated memory systems. He discusses episodic memory, suggesting it involves indexing temporal experiences within a learned statistical model of the world, likely involving interaction between cortical and hippocampal structures. This approach aims to build AI that can reason, simulate, and learn from experience in a more human-like manner.

BRAIN-COMPUTER INTERFACES AND THE NATURE OF INTELLIGENCE

Regarding Brain-Computer Interfaces (BCIs) and Neuralink, George views them as a promising field, particularly for therapeutic applications. He suggests that direct neural communication, while facing significant safety and adaptation challenges, might progress by the brain adapting to the interface rather than a perfect protocol being derived upfront. He also touches on consciousness and motivation, suggesting that mortality, a fundamental aspect of biological existence, may not be a necessary component for AI intelligence, as AI systems could potentially be copied or backed up, avoiding the finite lifespan that drives human urgency and goal-setting. The focus remains on understanding the machinery of the world to enable self-defined goals.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Books

●Concepts

●People Referenced

Common Questions

Yes, Dileep George argues that building a functional brain requires understanding its underlying theory and how its pieces fit together, rather than just simulating details like in the Blue Brain Project. Without a theory, debugging and predicting behavior is impossible.

Topics

AI & Machine Learning Neuroscience & the Brain Mindset & Self-Improvement Brain-inspired AI Computational Neuroscience Visual Perception Cognitive Architectures Language Understanding Deep Learning Limitations Episodic Memory Brain-computer Interfaces Philosophy Of AI Career Advice

Mentioned in this video

Companies

Numenta

A company co-founded by Dileep George with Jeff Hawkins and Donna Dubinsky, focusing on brain-inspired AI.

Paradromics

Another company working in the Brain-Computer Interface space, using a different electrode array.

Vicarious

A company co-founded by Dileep George, that builds brain-inspired AI with a focus on modeling the world and internalizing external actions.

DeepMind

A British artificial intelligence subsidiary of Alphabet Inc., mentioned in the context of AlphaZero and its self-play mechanisms.

Neuralink

Elon Musk's company focusing on developing brain-computer interfaces, discussed for its potential applications and challenges in safety and integration.

OpenAI

An AI research and deployment company, whose work on Rubik's Cube and GPT-2/3 models has drawn both attention and criticism.

Concepts

Hodgkin-Huxley Model

A mathematical model that describes how action potentials in neurons are initiated and propagated, used in the Blue Brain Project.

Probabilistic Graphical Models

A class of machine learning models where each node is a variable representing a concept, used for encoding knowledge and doing inference.

Hierarchical Temporal Memory

An early work by Dileep George, a brain-inspired model for machine intelligence.

Expectation-maximization (EM) algorithm

An iterative method to find maximum likelihood or maximum a posteriori (MAP) estimates of parameters in statistical models, mentioned as another weight adjustment algorithm in learning.

Markov chain

A mathematical model describing a sequence of possible events in which the probability of each event depends only on the state attained in the previous event, used in Shannon's model of English text.

Thalamus

A subcortical structure through which all cortical columns pass, hypothesized to combine with cortical neurons for inference computations.

Kanizsa Triangle

A visual illusion where a white equilateral triangle is perceived in an image, but it's not actually there, demonstrating the brain's ability to hallucinate edges from context using feedback connections.

Schema Networks

A type of cognitive learning system developed by Dileep George, which is a controllable dynamic system for learning concepts and eventually connecting to language.

graph neural networks

A class of neural networks for processing data that can be represented as graphs, showing a convergence of neural network and graphical model properties.

convolutional neural networks

A type of neural network with feed-forward processing cascades for feature detection and pooling, which Dileep George contrasts with RCN.

Brain-computer interface

A direct communication pathway between an enhanced or wired brain and an external device, discussed as a fascinating and promising research area.

Recursive Cortical Networks

A brain-inspired approach to computer vision developed by Dileep George, emphasizing feedback and lateral connections, and dynamic inference.

Transformer

A novel neural network architecture used in deep learning, particularly for natural language processing, which GPT-3 is based on.

Organizations

Redwood Neuroscience Institute

An institute where Jeff Hawkins organized workshops, including discussions on ideal undergraduate majors for understanding the brain.

Science (journal)

A highly-respected academic journal where Dileep George's paper on Recursive Cortical Networks was published.

Blue Brain Project

Henry Markram's project attempting to simulate a brain at a very detailed biophysical neuron level, which Dileep George argues is unlikely to work without a functional theory.

People

Krishna Shenoy

A professor at Stanford University, whose BCI talk inspired Dileep George's initial interest in brains.

Jeff Hawkins

Co-founder of Numenta and author of 'On Intelligence'.

Elon Musk

Visionary entrepreneur, behind companies like Neuralink and SpaceX, known for ambitious timelines.

Douglas Hofstadter

An American scholar of cognitive science and comparative literature, whose talks and books on AI (like 'The Mind's I') influenced Dileep George.

Daniel Dennett

An American philosopher, cognitive scientist, and writer, who co-authored 'The Mind's I'.

Judea Pearl

A computer scientist and philosopher, and Turing Award laureate, known for his work on causality and probabilistic reasoning.

Donna Dubinsky

Co-founder of Numenta.

Noam Chomsky

A linguist and political activist, whose concept of universal grammar is disagreed with by Dileep George in favor of grounded, pre-verbal concept understanding.

Wright brothers

Pioneers of aviation who invented and built the world's first successful airplane, whose observation of birds led to fundamental insights about flight control.

Claude Shannon

An American mathematician, electrical engineer, and cryptographer known as 'the father of information theory', whose work on statistical models of English text is mentioned.

Dileep George

A researcher at the intersection of neuroscience and artificial intelligence, co-founder of Vicarious and formerly co-founder of Numenta. He focuses on engineering intelligence inspired by the human brain.

Scott Phoenix

Co-founder of Vicarious.

Books

Probabilistic Reasoning in Intelligent Systems

A deep technical book by Judea Pearl, recommended for its philosophical observations sprinkled throughout its explanation of probabilistic graphical models.

On Intelligence

A book by Jeff Hawkins that outlines his theory of intelligence, discussed in the context of its high-level ideas being consistent with current brain models.

The Mind's I

A book by Douglas Hofstadter and Daniel Dennett, recommended by Dileep George.

Science Robotics

A scientific journal where Dileep George's paper on cognitive programs applied to manipulation robotics was published.

The Book of Why

A more accessible book by Judea Pearl on causality.

Software & Apps

GPT-3

A large language model by OpenAI, discussed for its text generation and coherence, but critiqued for its lack of a true world model and ability to run simulations.