AI summary by TwinMind
Can AI Match the Human Brain? | Surya Ganguli | TED
TED
17m 10s
Introduction and Context
- Strange new type of intelligence appeared on planet in last decade
- AI has remarkable capabilities but makes egregious errors humans never make
- Lacks deep logical reasoning, has mysterious surface of capabilities and fragilities
- Almost no understanding of how AI works
- Need deeper scientific understanding of intelligence in historical context of biological intelligence
Historical Context of Intelligence Evolution
- Last common ancestor of vertebrates lived 500 million years ago
- Evolution built brain, which in 500 years (Newton to Einstein) developed deep math/physics to understand universe from quarks to cosmology
- Recent AI advances require combining physics, math, neuroscience, psychology, computer science
- Engineering of intelligence has vastly outstripped ability to understand it
Five Critical Areas for AI Improvement
1. Data Efficiency Gap
- AI vastly more data-hungry than humans
- Language models trained on ~1 trillion words vs humans get only 100 million words
- Would take humans 24,000 years to read 1 trillion words
- Human DNA contains only 700 megabytes (600 million words) of evolutionary information
- Combined learning + evolution information minuscule compared to AI requirements
Scaling Laws Problem
- Current scaling laws have small exponents
- To reduce error slightly, need 10x more training data
- Unsustainable long-term approach
- Large random datasets incredibly redundant
- Next data point doesn't provide much new information
Solution: Nonredundant Datasets
- Theory and algorithms developed for carefully chosen data points
- Each point tells something new compared to all others
- Theoretically predicted and experimentally verified
- Bent bad power laws down to better exponentials
- Few more data points can reduce error vs 10x data requirement
- Used statistical physics equations
Machine Teaching Approach
- Current pretraining like teaching kids through next-word prediction on random internet chunks
- Would take kids 24,000 years to learn anything useful
- Humans teach algorithms (e.g., math) enabling immediate problem solving with less data
- Need to go beyond current training algorithms
- Transform machine learning into science of machine teaching
2. Energy Efficiency Gap
- Human brains consume only 20 watts (less than 100-watt light bulbs)
- Training large AI models consumes 10 million watts
- Talk of nuclear-powered billion-watt data centers
Root Cause: Digital Computation
- Relies on fast, reliable bit flips at every intermediate step
- Thermodynamics demands fast/reliable bit flips consume lots of energy
- Biology computes right answer just in time
- Uses intermediate steps as slow and unreliable as possible
- Biology doesn't rev engine more than needed
Biology's Computational Advantages
- Matches computation to physics better
- Example: Neurons directly add voltage inputs using Maxwell's electromagnetic laws
- Computers use complex energy-consuming transistor circuits for addition
- Biology matches computation to native physics of universe
Research Solutions
- Rethink entire technology stack from electrons to algorithms
- Match computational dynamics to physical dynamics
- Found fundamental limits on speed/accuracy given energy budget
- Solved for sensing computation (every neuron requirement)
- Found chemical computers achieving these limits
- Results looked like G-protein coupled receptors used by neurons
- Biology achieves efficiency close to fundamental physics limits
Predictive Energy Allocation
- Measured neural activity and ATP consumption across fly brain
- When neural activity increases, ATP goes up (not down as expected)
- ATP stays elevated long enough to power expected future neural activity
- Brain follows predictive energy allocation principle
- Predicts energy needs by location and timing
- Delivers right amount of energy at right location for right duration
3. Going Beyond Evolution
- Co-opt neural algorithms discovered by evolution
- Implement in quantum hardware evolution couldn't figure out
- Replace neurons with atoms (different firing states correspond to atomic energy levels)
- Replace synapses with photons (enable atom-to-atom communication)
Quantum Neuromorphic Computing
- Built quantum associative memory from atoms and photons
- Same memory system that won John Hopfield Nobel Prize in physics
- Quantum dynamics yield enhanced memory capacity, robustness, recall
- Built new quantum optimizers directly from photons
- Solve optimization problems in fundamentally new ways
- Opens entirely new field of quantum neuromorphic computing
4. Explainable AI for Brain Understanding
- AI builds accurate but complicated brain models
- Risk: replacing something we don't understand (brain) with something else we don't understand (complex model)
- Need conceptual understanding, not just handed model
Retina Research Example
- Built world's most accurate retina model
- Reproduced two decades of retina experiments
- Created digital twin of retina
Newton's First Law Violation Detection
- Demonstration: hand tracking experiment showing motion reversal
- Retina neurons fire if/only if Newton's first law violated
- Model reproduced this behavior
- Developed explainable AI methods to carve out essential subcircuits
- Explained how subcircuits work for two decades of experiments
- Model makes new predictions
Stanford Digital Twin Project
- Building digital twin of entire primate visual system
- Use explainable AI to understand how it works
- New pathway to accelerate neuroscience discovery
5. Melding Minds and Machines
- Bidirectional communication between brains and machines
- Process: record from brain → build digital twin → learn neural control patterns → write patterns back to brain
- Learn language of brain and speak directly back to it
Mouse Mind Reading/Control Experiments
- Used AI to read mouse mind
- Top row: actual images shown to mouse
- Bottom row: images decoded from mouse brain
- Decoded images show world as mouse actually sees it (lower resolution due to mouse visual limits)
- Can write neural activity patterns into mouse brain
- Make mouse hallucinate any desired percept
- Control mouse visual perception by controlling only 20 specific neurons
- Achieved reliable hallucination control
Applications
- Limitless possibilities for bidirectional brain-machine communication
- Understand, cure, and augment the brain
