A Theory of Cerebral Cortex

1

• A Theory of Cerebral Cortex
• (or, How Your Brain Works)
• Andrew Smith (CSE)

2
Outline

• Questions
• Preliminaries
• Feature Attractor Networks
• Antecedent Support Networks
• Attractive properties of the theory / Conclusions

3
Questions (to be answered!)

• What is cortical knowledge and how is it stored?
• How is it used to carry out thinking?
• How is it integrated with sensory input and
  motor output?

4
Preliminaries

• Thinking is a symbolic process.
• Thinking relies only on classical mechanics.
  (Unlike the Penrose/Hameroff model.)
• Thinking is not a mathematically grounded
  reasoning process, rather confabulation!

5
Feature Attractor Neuronal Networks
Each Feature Attractor Network Implements one
Column of Tokens
cortical region (one of about 120,000)
cerebral cortex
human cortical surface area ? 240,000 mm2
a feature attractor network
paired thalamic region
bidirectional connections
thalamus
An object (sensory, abstract, etc.) or action
(movement process, thought process, etc.) is
represented by a collection of feature attractor
tokens, each expressing a single token (node)
from its lexicon.
Each region encompasses a cortical surface area
of roughly 2 mm2 and possesses a total of about
200,000 neurons.
6
Feature Attractor Networks

• Each network has a lexicon of random (!) tokens,
  sparsely encoded each token has 100s of neurons
  on at a time, out of 50,000. This lexicon is
  fixed very early in life and never changes.
• The function of the network is to change the
  pattern of activation within a particular region
  so that it expresses the token in its lexicon
  closest to the original pattern of activation.
  (aka vector quantizers)
• The Feature Attractor Networks are extremely
  robust to noise/partial tokens.
• - A region can start out with 10 of a
  particular token and
• within one iteration, express the complete
  token.
• - A region can start out expressing many
  (100s) of partial
• tokens and within one iteration, express just
  one token that was
• most complete. (more on this later)
• Now we have 120,000 powerful pattern
  recognizers, lets wire them up

7
Antecedent Support Networks (ASNs)

• The role of the ASN is to do the thinking.
• - If several active tokens have strong links to
  an inactive token, the ASN will activate that
  token
• (e.g. smoke heat -gt fire).
• - Learning occurs when the ASN increases the
  link weight between two tokens.
• Short term memory Which tokens are currently
  active
• Long term memory The link strengths between
  tokens

8
Antecedent Support Neuronal Network
Implementation Randomness to the rescue!
Axons from neuron of token i send their
collaterals randomly to millions of neurons in
the local area. Of these, a few thousand
transponder neurons just happen to receive
sufficient input from i to become active. Of
those, a few hundred just happen to send axons to
neurons belonging to token j on the target
region, activating (part of) token j.
The wiring of transponder neurons (pyramidal
neurons) is also fixed at a very early age.
transponder neurons
these are the synapses that are strengthened
target region token j
cerebral cortex
source region token i
9
Input / Output

• Input Sensory neurons connect to token neurons
  (layers III and IV), just like transponder
  neurons.
• Output Motor neurons can receive their inputs
  from the token neurons, just like transponder
  neurons.

10
Attractive features (no pun intended)

• The Hecht-Nielsen model shows
• - how neurons can grow randomly and become
  organized.
• - that a large range of synaptic weights is not
  necessary.
• - how you can get a song stuck in your head.
  (Youre unable to reset regions of your cortex.
  One bar evokes the next)
• - a model that can be viewed as implementing
  Paul Churchlands semantic maps from last
  lecture of CogSci 200. (IMHO)
• A simulation of this has solved the classic
  cocktail-party problem.

11
Conclusions

• All knowledge comes from creating associations
  between experiences.
• - Aristotle
• Within 12 to 36 months, this theory will
  revolutionize Artificial Intelligence.
• - Hecht-Nielsen
• (as of last week)