Thousand Brain Theory of Intelligence · jaysen naidoo's public-Txt

The Thousand Brains Theory

Through his company Numenta and the Thousand Brains Project, Hawkins applies neuroscience to machine learning:

Cortical Columns: The neocortex has roughly 150,000 repeating units called cortical columns, each acting as an independent learning machine.
Voting Consensus: These columns build parallel models of objects and vote together to create a unified perception.
Reference Frames: The brain uses grid cells as internal maps to anchor what things feel, look, or sound like during movement.

Numenta Jeff Hawkins
extends the Hierarchical Temporal Memory model for intelligence in the Neocortex
Reference Frames
Spatial referencing
- based on evolved for function of locating yourself in space - Grid Neurons
- Neural Recycling – idea from How We Learn - Stanislas Dehaene
- could be multidimensional - abstract conceptual space
Sparse Representations - Sparsely Distributed Representations
Learning and Synaptogenesis
- Long Term memory needs new synapses created
- Short term memory using existing but inactive synapses in hippocampus - Silent synapses . they can flip quickly from inactive to active
Why Neurons have 100s of Synapses #paper https://numenta.com/neuroscience-research/research-publications/papers/why-neurons-have-thousands-of-synapses-theory-of-sequence-memory-in-neocortex/
Neurons / Synapses /
- Real Neurons - from 5 to 30000 synapses
- 5% of synapses are close by - can affect the soma the cell body
- 95% are distant
- Distant synapses when they fire create dendritic spikes . which raises voltage of the cell (potentiation but not enough to make neuron fire.
- instead, distant synaptic firings decreases current neuron’s firing time. which is important to beat inhibitory process from neighbouring reference frames. see Inhibitory Neurons
- the neuron is a predictive engine on its own