Summary

Numenta founder Jeff Hawkins describes a new theory of intelligence and how we can leverage these mechanisms to advance machine learning in the pursuit of AGI. The neocortex holds thousands of models of “objects” (what) and reference frames (where) developed through sensory-motor and Hebian learning. Prediction and learning is done by “voting” between complex combinations of distributed, hierarchical, complementary models. Hawkins argues that intelligence is about how a machine learns and not how it performs tasks, thus multiple technological advancements are still needed to match the capabilities of the brain via AGI. As he sees it, we have a choice between 1) the creation and the dissemination of knowledge and 2) the copying and dissemination of genes. Consciousness needs to be vitally preserved as it is only entity aware of our universe.

“We have a profound choice to make: it is a choice between favoring the old brain or the new brain. More specifically, do we want our future to be driven by the processes that got us here - namely natural selection, competition, and the drive of selfish genes - or do we want our future to be driven by intelligence and the desire to understand the world?”

Key Takeaways

• Neocortex: New brain comprising 80% of brain, ~150,000 copied general purpose cortical columns
  • Sensory-motor, Hebian learning through thousands of predictions (unsupervised)
  • Inference through “voting” between complex combinations of distributed, hierarchical, complementary models
• Multiple technological advancements are still needed to match the capabilities of the brain via AGI
  • Learns continuously, movement, voting via many models, general purpose reference frames
  • Consciousness: (aware) requires forming stream of memories of actions, thoughts
  • Neocortex learns models without goals or desires, old brain attaches goals + emotion
    • AI should have built-in, learned behaviors akin to habits / muscle memory in old brain
• Choice between creation + dissemination of knowledge and copying + dissemination of genes
  • Curiosity (exploration, space travel) is just selfish genes evolving to further proliferate
  • Should focus scientific, educational efforts on the brain, intelligence, knowledge
  • Consciousness is only thing aware of universes’ age, laws, etc. - need to preserve
• Better to broadcast, seek help rather than try and keep secret (eg. entrepreneurs)
• Origin of false beliefs:
  • Cannot directly experience
  • Ignoring contrary evidence: Only way to validate model is to actively seek out contrary evidence, eg. scientific method
  • Viral model spread: proliferates through society (eg. educating children)

Notes

1: A new understanding of the brain

1: Old brain, new brain

• Old brain: (30%, many parts) that controls bodily functions, emotions, hormones etc.
• New brain: (70%, all neocortex) complex thoughts, against gene evolution
• Brain learns by forming models of world
• Classically 6 neural “layers” of neocortex
• Dozens of different neuron cell types
• Brain mostly homogeneous

2: Vernon Mountcastle’s big idea

• Neocortex: (small part of original brain) evolved by copying same basic structure
• Like Darwin’s single algo for evolution, Mountcasle’s single algo for intelligence
• Cortical column: 2.5mm^3 unit of neocortex with grouped response to stimulus
• ^ subdivided into 100s of mini-columns
• ~150,000 columns in brain
• Brain has general-purpose method for learning

3: A model of the world in your head

• “Prediction is a ubiquitous function of the neocortex”
• Neocortex makes thousands of small predictions and learns based on results
• Brain builds models (eg. object location, looks, use, sound) to predict inputs
• Sensory-motor learning: learning by moving and sensing / predicting sensations
• Hebian learning: neural connections strengthen / weaken with use (associative)
  • Foundation for unsupervised learning

4: The brain reveals its secrets

• Neurons take dendrite inputs and activate in certain circumstances
• 90% of dendrites are for prediction, not communication
• Reference frame: framing any concept
  • Eg. Location, orientation in 3D space
• Each cortical column makes predictions relative to a reference frame

5: Maps in the brain

• Place cells / grid cells: create a map of the environment in old brain
• Work together to learn maps on environments (spacial)
• Also in charge of orientation
• Similar cellular apparatus in neocortex

6: Concepts, language, high-level thinking

• Reference frames are present everywhere in the neocortex
• Reference frames are used to map / model everything, not just physical objects
  • Physical objects (what), space around body (where), concepts (non-sensory)
• Same cortical column can play different roles given different reference frames
• “Random” thoughts depend on which way we move through mental reference frame
• Language handled by slightly different looking region, assume it works the same
• Language: nested, recursive qualities
• “Reference frames provide the substrate for learning the structure of the world - where things are and how they move and change”

7: The thousand brains theory of intelligence

• Hierarchy of features theory: flowchart from sensors through feature detectors, building higher levels of understanding
• No concrete store of knowledge in brain, item distributed b/w complimentary models
• “Voting” among different columns of brain for collective reasoning
• “Neocortex uses hierarchy to assemble objects into more complex objects”
• Neocortex still largely not understood, but basic framework of reference frames should for basis for future understanding

2: Machine intelligence

8: Why there is no “I” in AI

• Specialized vs universal Turing machines akin to today’s AI vs AGI
• Market forces will force AGI discovery like they did with Turing machines
• 4 baseline elements for general intelligence:
  • Learns continuously (flexible)
  • Movement (physical + intellectual)
  • Many models (voting)
  • General purpose reference frames
• Intelligence is about how a machine learns, not how it performs tasks

9: When machines are conscious

• Consciousness: (aware) requires forming stream of memories of actions, thoughts
• Qualia: how sensory inputs are perceived
  • Internal aspect of consciousness
  • Can’t confirm perceived the same across different brains
• Not all qualia is learned (eg. pain is innate)
• Fear of death is phenomenon of old brain and not required for consciousness
  • Can ethically turn off conscious AI

10: The future of machine intelligence

• Neocortex learns models without goals or desires, old brain attaches goals + emotion
• Embodiment: having sensors with ability to move them to learn models
• AI should have built-in, learned behaviors akin to habits / muscle memory in old brain
• Must have goals and desires
• Like advent of internet, hard to imagine full scale of use cases for strong AI

11: The existential risks of machine intelligence

• Hawkins argues that AGI risks overblown:
  • Intelligence explosion (learning is slow)
  • Can’t best humans in all tasks (moving bar)
  • Goal misalignment (assumes capability, no subsequent goal updates)
• Right now, AI only excels in static tasks
• Replications, motivations, intelligence

3: Human intelligence

12: False beliefs

• Brain experiences simulation of earth based on electric spikes from sensory systems
• Viral model: a model which proliferates through society (eg. educating children)
• Viral false beliefs: (eg. Bible)
• Only way to validate model: actively seek out contrary evidence (scientific method)

13: The existential risks of human intelligence

• We can overcome genetic tendencies (eg. overpopulation)
• Origin of false beliefs:
  • Cannot directly experience
  • Ignoring contrary evidence
  • Viral spread

14: Merging brains and machines

• Uploading brain isn’t trivial: exact replica would need to adapt sensors, motors etc.
• “Copying yourself is a fork in the road, not an extension of it”
• Having children is similar to mind uploading

15: Estate planning for humanity

• Knowledge: what we have learned about the world
• As humankind, we should plan our legacy and how we can communicate our knowledge to future intelligence, both terrestrial and ET
• Better to broadcast, seek help rather than try and keep secret (eg. entrepreneurs)

16: Genes versus knowledge

• Should become an interplanetary species by preparing Mars with AI robots, colonizing
• “We have a profound choice to make: it is a choice between favoring the old brain or the new brain. More specifically, do we want our future to be driven by the processes that got us here - namely natural selection, competition, and the drive of selfish genes - or do we want our future to be driven by intelligence and the desire to understand the world?”
• Choice between creation + dissemination of knowledge and copying + dissemination of genes
• Gene editing allows us to control our genes, could be essential for survival of species
• Curiosity (exploration, space travel) is just selfish genes evolving to further proliferate
• Should focus scientific, educational efforts on the brain, intelligence, knowledge
• Consciousness is only thing aware of universes’ age, laws, etc. - need to preserve