Transcript Can a Machine Be Conscious?

A New Theory of Neocortex and Its
Implications for Machine Intelligence
TTI/Vanguard, All that Data
February 9, 2005
Jeff Hawkins
Director
The Redwood Neuroscience Institute
Intelligence Paradigms
Artificial Intelligence (AI)
- ignores biology
- computer programs
- emulate human behavior
1940s - 1980s
Neural Networks
- mostly ignores biology
- networks of “neurons”
- classify spatial patterns
1970s - 1990s
Intelligence Paradigms
Artificial Intelligence (AI)
- ignores biology
- computer programs
- emulate human behavior
1940s - 1980s
Neural Networks
- mostly ignores biology
- networks of “neurons”
- classify spatial patterns
1970s - 1990s
“Real Intelligence”
2005 –
- biologically derived
- hierarchical temporal memory
- pattern prediction
Hierarchical Temporal Memories (HTMs)
A Fundamental technology
Automatically discover causes in complex systems
Predict future behavior of complex systems
Can build super-human intelligence (not C3PO)
- faster
- more memory
- novel senses
Agenda
Introduction to neocortex
What does the neocortex do?
How does it do it?
Can we express this mathematically?
How do we build it?
What problems can be solved?
Agenda
Introduction to neocortex
What does the neocortex do?
How does it do it?
Can we express this mathematically?
How do we build it?
What problems can be solved?
Agenda
Introduction to neocortex
What does the neocortex do?
How does it do it?
Can we express this mathematically?
How do we build it?
What problems can be solved?
1) The neocortex is a memory system.
2) Through exposure, it builds a model the world.
3) The neocortical memory model predicts future events
by analogy to past events.
Reptilian brain
Reptilian brain
Behavior
Sophisticated
senses
Mammalian brain
Neocortex
Reptilian brain
Behavior
Sophisticated
senses
Human brain
Neocortex
Reptilian brain
Complex
behavior
Sophisticated
senses
Agenda
Introduction to neocortex
What does the neocortex do?
How does it do it?
Can we express this mathematically?
How do we build it?
What problems can be solved?
Hierarchical connectivity
touch
motor
audition
vision
spatially
invariant
slow
changing
“objects”
spatially
specific
fast
changing
“features”
“details”
Prediction
touch
motor
audition
vision
Prediction across senses
touch
motor
audition
vision
Sensory/motor integration
touch
motor
audition
vision
touch
motor
audition
vision
touch
motor
audition
vision
What does each region do?
?
touch
motor
audition
vision
What does each region do?
Every region:
1) Stores sequences
2) Passes sequence “name” up
3) Predicts next element
4) Converts invariant prediction
into specific prediction
5) Passes specific prediction “down”
touch
motor
audition
vision
Hierarchical cortex captures hierarchical structure of world
- sequences of sequences - structure within structure
Unanticipated events
rise up the hierarchy
until some region can
interpret it.
Hippocampus is at the top.
Novel inputs that cannot be explained as part of
known structure automatically rise to the top.
HC
Unanticipated events
rise up the hierarchy
until some region can
interpret it.
Hierarchical Temporal Memories
Can Explain Many Psychological Phenomena
- Creativity, Intuition, Prejudice
- Thought
- Consciousness
- Learning
How does a region work - biology
Every region:
1) Stores sequences
2) Passes sequence “name” up
3) Predicts next element
4) Converts invariant prediction
into specific prediction
5) Passes specific prediction “down”
Agenda
Introduction to neocortex
What does the neocortex do?
How does it do it?
Can we express this mathematically?
How do we build it?
What problems can be solved?
All inputs and outputs from a memory region are
probability distributions
Higher regions
Lower regions
Learning
Higher regions
C
P(S|C)
SA(xt,xt+1,...)
SB(xt,xt+1,...)
X
Lower regions
C = causes or context
S = sequences
X = input
Recognition without context
Higher regions
P(C)
P(S|C)
SA(xt,xt+1,...)
SB(xt,xt+1,...)
X
Lower regions
Recognition with context can lead to new interpretation
Higher regions
C1
C1
P(S|C)
SA(xt,xt+1,...)
SB(xt,xt+1,...)
X
Lower regions
Passing a belief down the hierarchy
Higher regions
C
C
P(S|C)
SA(xt,xt+1,...)
SB(xt,xt+1,...)
Xt
f ( Xt, P(S|C) )
Lower regions
Predicting the future
Higher regions
C
C
P(S|C)
SA(xt,xt+1,...)
SB(xt,xt+1,...)
Xt
f ( Xt+1, P(S|C) )
Lower regions
Belief Propagation can determine most likely causes of input
in a hierarchy of conditional probabilities
P(X)
P(Y1|X)
P(Z1|Y1)
P(Z2|Y1)
P(Y2|X)
P(Z3|Y1)
P(Z4|Y1)
System Architecture
Level 3
Level 2
Level 1
4 pixels
Recognition : Examples
Correctly Recognized
“Incorrectly” recognized
Correctly Recognized Test Cases
Prediction/Filling-in : Example1
Prediction/Filling-in : Example2
What’s new?
Hierarchical
Neocognitron
HMax
Seemore, Visnet
Sequence memory
auto-associative memories
synfire chains
Prediction/feedback
HMMs
ART
Sensory/motor integration
Biologically derived/constrained/testable
Agenda
Introduction to neocortex
What does the neocortex do?
How does it do it?
Can we express this mathematically?
How do we build it?
What problems can be solved?
Hierarchical Temporal Memories (HTMs)
A Fundamental technology
Automatically discover causes in complex systems
Predict future behavior of complex systems
Can build super-human intelligence (not C3PO)
- faster
- more memory
- novel senses
What problems can be solved with HTMs?
Traditional AI applications
- Vision
- Language
- Robotics
Novel modeling applications
- markets
- weather
- demographics
- protein folding
- gene interaction
- mathematics
- physics
www.OnIntelligence.org
www.stanford.edu/~dil/invariance/
Thank ---
Learning sequences
L5/matrix thalamus/L1 auto-associative loop
Creating a sequence “name”
Turning an invariant prediction into a specific
prediction