Editor's note: Jeff Hawkins is an American inventor, computer scientist, and neuroscientist. He led the development of Palm and Treo, and is the founder of Palm, Numenta, and Handspring. In this article, Jeff shared his unique insights on machine intelligence from the three aspects of biological composition, functional composition, and diversification of intelligent machines.
1) Intelligent biological components
2) Intelligent functional components
3) Diversification of smart machines
Intelligent biological components
The importance of Neocortex (new (cerebral) cortex)
Neocortex is located between the hippocampus (which is considered part of the emotional and memory center in the brain) and the rest of the brain. It occupies 75% of the entire brain volume and helps people to better understand the outside world.
Hierarchical characterization of structures
1) The brain regions of different species and patterns have significant similarities (thus, all regions of the brain perform the same kinetic energy)
2) Different levels of brain area hierarchy appear between different species (thus, the hierarchy diagram is not critical, it only represents a design parameter)
Each area of ​​the brain has the following capabilities at the same time :
1) Identify sensory sequences (eg oral language, music, visual actions) (input)
2) Identify sensory-motion sequences (eg, limb movement, eye movement) (input)
3) Generate Action Sequence (Output)
Each area performs exactly the same function as the entire hierarchy.
Reasoning: Sequence memory is an important function of every brain area.
Pyramidal neurons
Each pyramidal neuron can identify hundreds or thousands of independent patterns.
The pattern recognized at the basal dendrites will depolarize the cells but will not generate action potentials.
Hypothesis: Depolarization is a prediction that a depolarized cell will be activated soon and will have an inhibitory effect on nearby cells.
Represents higher order series ABCD vs. XBCY
In real data, the sequence composition is very complicated.
Apical synaptic prediction sequence
Learning through synapses (this is a very efficient way to learn)
(Note: A single synapse has a strong randomness)
Dendrite dendrites
Axon Axon
Synapse permanence synapses are permanent
Synapse weight synaptic weight
Hierarchical Temporal Memory Based Sequential Memory
- Continuous learning
- No batch training
- Adjust as the mode changes
- Robust cell death
Intelligent functional components
Hawkins list of functional components of intelligence(This list is subjective)
1) Neuron networks with learning and recalling sequence capabilities
- Continuous learning, no batch training
- Make multiple predictions at the same time
- Robustness
Hierarchical real-time memory: active dendrites, synapse formation, no cones
2) Each area of ​​the brain uses sequence memory for :
- Sensory reasoning
- Sense - Action Reasoning
- Action generation
(The above three uses are the basic conditions that smart machines and even human intelligence should have)
3) Hierarchy of brain regions
- The number of areas
- The size of the area
- Connectivity map of brain area
4) Body
--sensor
- Embedded behavior
- Emotion/motivation
——Scene Memory/Space Memory
Diversification of smart machines
Jeff believes that one day we will be able to develop a machine with a super mathematician or physicist talent. We need to construct a machine with a hierarchical feature of the human brain. This machine will have the mathematician behavior pattern and be able to perform mathematics. The function of the home, however, is to achieve machine intelligence, and we still have a long way to go on the road to research.
Via Jeff Hawkins
Article video link
This article was compiled by Lei Feng Network (search "Lei Feng Net" public number) , and refused to reprint without permission!
Rice Polisher Polishing Machine,Rice Polisher Emery Roller,Rice Polisher,Whitener Rice Mill
HuangShi Stee Metal Material Technology Exploitation Co.,Ltd , https://www.wosenstechnology.com