Transcript Design of Intelligent Machines Heidi 2005

Design of Self-Organizing Learning
Array for Intelligent Machines
Janusz Starzyk
School of Electrical Engineering
and Computer Science
Heidi Meeting June 3 2005
Motivation:
How a new understanding of the brain will lead
to the creation of truly intelligent machines
from J. Hawkins “On Intelligence” 1
EE141
Elements of Intelligence










Abstract thinking and action planning
Capacity to learn and memorize useful things
Spatio-temporal memories
Ability to talk and communicate
Intuition and creativity
Consciousness
Emotions and understanding others
Surviving in complex environment and adaptation
Perception
Motor skills in relation to sensing and anticipation
2
EE141
Problems of Classical AI
 Lack
of robustness and generalization
 No real-time processing
 Central processing of information by a
single processor
 No natural interface to environment
 No self-organization
 Need to write software
3
EE141
Intelligent Behavior
Emergent from interaction with environment
 Based on large number of sparsely connected
neurons
 Asynchronous
 Self-timed
 Interact with environment through sensorymotor system
 Value driven
 Adaptive

4
EE141
Design principles of intelligent systems
from Rolf Pfeifer “Understanding of Intelligence”
Design principles
synthetic methodology
time perspectives
emergence
diversity/compliance
frame-of-reference
Agent design
complete agent principle
cheap design
ecological balance
redundancy principle
parallel, loosely coupled
processes
sensory-motor coordination
value principle
5
EE141
The principle of “cheap design”

intelligent agents: “cheap”
 exploitation of ecological
niche
 economical (but redundant)
 exploitation of specific
physical properties of
interaction with real world
6
EE141
Principle of “ecological balance”

balance / task distribution
between
 morphology
 neuronal processing (nervous
system)
 materials
 environment

balance in complexity
 given task environment
 match in complexity of sensory,
motor, and neural system
7
EE141
The redundancy principle



redundancy prerequisite for
adaptive behavior
partial overlap of
functionality in different
subsystems
sensory systems: different
physical processes with
“information overlap”
8
EE141
Generation of sensory stimulation
through interaction with environment
multiple modalities
 constraints from
morphology and
materials
 generation of
correlations through
physical process
 basis for crossmodal associations

9
EE141
The principle of sensory-motor
coordination
Holk Cruse



self-structuring of
sensory data through
interaction with
environment
physical process —
not „computational“
prerequisite for
learning
•no central control
•only local
neuronal
communication
•global
communication
through
environment
neuronal
connections
10
EE141
The principle of parallel, loosely
coupled processes
Intelligent behavior emergent
from agent-environment
interaction
 Large number of parallel,
loosely coupled processes
 Asynchronous
 Coordinated through agent’s
–sensory-motor system
–neural system
–interaction with environment

11
EE141
Neuron Structure and SelfOrganizing Principles
Human
Brain
at Birth
EE141
6 Years Old
14 Years
Old
12
12
Neuron Structure and SelfOrganizing Principles (Cont’d)
13
EE141
Motor cortex
Somatosensory cortex
Sensory associative
cortex
Pars
opercularis
Visual associative
cortex
Broca’s
area
Visual
cortex
Primary
Auditory cortex
Wernicke’s
area
Brain Organization
While we learn
its functions
can we emulate
its operation?
14
EE141
Minicolumn Organization and
Self Organizing Learning Arrays
V. Mountcastle argues that all regions of the
brain perform the same algorithm
 SOLAR combines many groups of neurons
(minicolumns) in a pseudorandom way
 Each microcolumn has the same structure
 Thus it performs the same computational
algorithm satisfying Mountcastle’s principle


VB Mountcastle (2003). Introduction [to a special issue of Cerebral
Cortex on columns]. Cerebral Cortex, 13, 2-4.
15
EE141
Cortical Minicolumns
“The basic unit of cortical operation is
the minicolumn … It contains of the
order of 80-100 neurons except in
the primate striate cortex, where
the number is more than doubled.
The minicolumn measures of the
order of 40-50 m in transverse
diameter, separated from adjacent
minicolumns by vertical, cellsparse zones … The minicolumn is
produced by the iterative division
of a small number of progenitor
cells in the neuroepithelium.”
(Mountcastle, p. 2)
Stain of cortex in planum temporale.
16
EE141
Groupping of Minicolumns
Groupings of minicolumns seem to form the
physiologically observed functional columns. Best
known example is orientation columns in V1.
They are significantly bigger than minicolumns, typically
around 0.3-0.5 mm and have 4000-8000 neurons
Mountcastle’s summation:
“Cortical columns are formed by the binding together of
many minicolumns by common input and short range
horizontal connections. … The number of
minicolumns per column varies … between 50 and
80. Long range intracortical projections link columns
with similar functional properties.” (p. 3)
17
EE141
Sparse Connectivity
The brain is sparsely connected.
(Unlike most neural nets.)
A neuron in cortex may have on the order of 100,000 synapses.
There are more than 1010 neurons in the brain. Fractional
connectivity is very low: 0.001%.
Implications:
 Connections are expensive biologically since they take up
space, use energy, and are hard to wire up correctly.
 Therefore, connections are valuable.
 The pattern of connection is under tight control.
 Short local connections are cheaper than long ones.
Our approximation makes extensive use of local connections for
computation.
18
EE141
Introducing Self-Organizing
Learning Array SOLAR
SOLAR is a regular array of identical processing
cells, connected to programmable routing
channels.
Each cell in the array has ability to self-organize by
adapting its functionality in response to
information contained in its input signals.
Cells choose their input signals from the adjacent
routing channels and send their output signals
to the routing channels.
Processing cells can be structured to implement
minicolumns
19
EE141
SOLAR Hardware Architecture
20
EE141
SOLAR Routing Scheme
21
EE141
PCB SOLAR
22
EE141
System SOLAR
23
EE141
Wiring in SOLAR
Initial wiring and final wiring selection for credit card
approval problem
24
EE141
SOLAR Classification Results
25
EE141
Associative SOLAR
26
EE141
Associations made in SOLAR
27
EE141
Brain Structure with Value System Properties







Interacts with environment through sensors and
actuators
Uses distributed processing in sparsely connected
neurons organized in minicolumns
Uses spatio-temporal associative learning
Uses feedback for input prediction and screening
input information for novelty
Develops an internal value system to evaluate its
state in environment using reinforcement learning
Plans output actions for each input to maximize the
internal state value in relation to environment
Uses redundant structures of sparsely connected
processing elements
28
EE141
Possible Minicolumn Organization
Value System
Understanding
Action
Planning
Expectation
Improvement
Detection
Comparison
Inhibition
Novelty
Detection
Anticipated Response
Reinf.
Signal
Sensors
EE141
Sensory
Inputs
Motor
Outputs
Actuators
29
Postulates for Minicolumn Organization








EE141
Learning should be restricted to unexpected situation
or reward
Anticipated response should have expected value
Novelty detection should also apply to the value
system
Need mechanism to improve and compare the value
Anticipated response block should learn the response
that improves the value
A RL optimization mechanism may be used to learn
the optimum response for a given value system and
sensory input
Random perturbation should be applied to the
optimum response to explore possible states and
learn their the value
New situation will result in new value and WTA will
30
chose the winner
Minicolumn Selective Processing







Sensory inputs are represented by more and more
abstract features in the sensory inputs hierarchy
Possible implementation is to use winner takes all or
Hebbian circuits to select the best match
“Sameness principle” of the observed objects to
detect and learn feature invariances
Time overlap of feature neuron activation to store
temporal sequences
Random wiring may be used to preselect sensory
features
Uses feedback for input prediction and screening
input information for novelty
Uses redundant structures of sparsely connected
processing elements
31
EE141
Minicolumn Organization
superneuron
Value
Positive
Reinforcement
Negative
Reinforcement
Sensory
Sensory
Inputs
Motor
Motor
Outputs
32
EE141
Minicolumn Organization

Sensory neurons are primarily responsible for providing information
about environment
 They receive inputs from sensors or other sensory neurons on lower level
 They interact with motor neurons to represent action and state of
environment
 They provide an input to reinforcement neurons
 They help to activate motor neurons

Motor neurons are primarily responsible for activation of motor functions
 They are activated by reinforcement neurons with the help from sensory
neurons
 They activate actuators or provide an input to lower level motor neurons
 They provide an input to sensory neurons

Reinforcement neurons are primarily responsible for building the
internal value system
 They receive inputs from reinforcement learning sensors or other
reinforcement neurons on lower level
 They receive inputs from sensory neurons
 They provide an input to motor neurons
 They help to activate sensory neurons
33
EE141
Sensory Neurons Functions
Sensory neurons
Represent inputs from environment by
Responding to activation from lower level
(summation)
Selecting most likely scenario (WTA)
WTA
Interact with motor functions by
Responding to activation from motor outputs
(summation)
Anticipate inputs and screen for novelty by
Correlation to sensory inputs from higher level
Inhibition of outputs to higher level
Select useful information by
Correlating its outputs with reinforcement neurons
Identify invariances by
Making spatio-temporal associations between
neighbor sensory neurons
WTA
WTA
34
EE141