Transcript 1 - UPM ASLab

Check if film is connected
Load NRM via Birm
Load nrmp6 and use guildford config aks2000 training and run naturally
'Digital Neuromodelling Based on the Architecture of the Brain: Basics and
Applications'
Digital neuromodelling is aimed at understanding the architecture of the brain. I
shall concentrate specifically on visual awareness. The principles of a digital
neuromon will be described and examples given of simple dynamic modules made of
such neurons . A software modelling tool (Neural Representation Modeller) will be
introduced and provided for the students to test on their own computers.
Examples of work done in this area will be given: anaesthesia;planning; visual
working memory; imagination mechanisms; visual deficits in Parkinson's disease and
explortory robotics.
The scope of the discussion will then be broadened to suggest an axiomatic
approach to the understanding of neural mechanisms responsible for being
conscious.
Turin Workshop on Machine
Consciousness
TUTORIAL
Digital Neuromodelling
Techniques
Igor Aleksander FREng
Leverhulme Fellow
Emeritus Professor of Neural Systems Engineering
DEPARTMENT OF
ELECTRICAL AND ELECTRONIC ENGINEERING
Imperial College, London
Visiting Research Fellow, U of Sussex
A Robot in our lab:
Driven by a neuromodel
rather than an AI program
The issue for today
How does neuromodelling differ
from AI programs and neural
networks?
What can it achieve?
We start with:
Who was the first person ever
to write an AI program?
Who is he?
Claude Shannon
of the
Bell Telephone Laboratories
“Programming a computer for playing chess”
Philosophical Magazine, pp 256-275, 1950
Shannon’s * algorithm
Each board
position has a value
ME
1, -----5-------10
YOU
ME
7
A stacking problem
?
The search space contains 85 possibilities
among which the computer needs to search.
A Search Process
The Brain:
•Works better in some areas.
•Is not pre-programmed but learns and
adapts.
•It has an architecture from which its
activity ‘emerges’ and which has
evolved.
Can it be modelled in these terms?
Where do the models start?
Neuron: The building brick of
the brain (one of 1011)
The Classical 1943 McCulloch and Pitts Neuron Model
What were they modelling?
THIS IS NOT WHAT WE
What are the elements of this model?
DO IN DIGITAL
x1
W
NEUROMODELLING
x2
W
Xj Wj

x
W
1
2
3
j
3
T
Wn
xn
F=1 iff

j
Xj Wj > T
F
The Digital Model of a Neuron
•Needs to learn and generalise
•Needs to make efficient use of
memory (10 million neurons)
•Needs to be fast (10 million
neurons updated in 1/4 Sec.)
A Digital neuron training 1
Input array
1
1
3
2
2
0
0
1
0
0
2
1
3
0
4
0
5
0
6
1
7
0
2
4
6
Train
4
5
State array
00100010 5
00012
5
Neuron
5
A Digital neuron training 2
Input array
Train
0
1
0
1
2
0
3
1
4
0
5
1
6
0
7
0
2
00100010 5
01010100 3
00012
Neuron
State array
3
Recall associated memory
Input array
State array
0
0
1
0
2
0
3
1
4
0
2
5
1
6
0
7
00100010 5
01010100 3
00012
vv . . v . . v
v . vvvvvv
Neuron
0
Recall
3
So, we have a picture of a neuron
but how do we test neural systems
on a computer?
We use a software authoring kit:
NRM: Neural Representation
Modeller
Email: [email protected]
How does a digital neuron work?
?
Colour recognition
1 neuron, 10,000 connections
(brain 15,000 connections)
How does a single layer of
digital neurons work?
name
3200 neurons, 20,480 connections
(Brain 100 billion neurons 1000 trillion
connections)
How do we go from a single
neuron to DEPICTIONS and
MEMORY in some parts of
the brain?
Communicating neurons: memory
and imagination
Using NRM for
Brain Modelling
Macaque monkey brain
Connections between different visual areas
Visual Awareness:
Basic Mechanisms
Macaque monkey brain
Visual areas shown in colour
muscle
eye
sc
Early, unconscious vision
Eye-centered
deconstruction
Conscious vision
World-centered
reconstruction:
perception, visual
memory
Neural Net
Superior
Colliculus
&
Motor Areas
Position
Gaze information
j
Extrastriate areas
(Gaze Locked)
Visual World
Depiction
Foveal eye
Neural Net
WORD
INPUT
WORD
DEPICTION
Visual Working Memory
(learned depictions)
Neural Net
Neural Net
UTTERANCE
Neural Representation Modeller
Visual Awareness
Area
Parkinson’s Disease
Visual
Consciousness
Deficits
A TYPICAL TASK
PARKINSON’S & VISION
)
BASAL G.
Eye Pos
Eye
Musc
Eye Mov
F
SC
Motor
F
PeriFov
PVCx
Gaze Lock Control
F
Fov
R.topic
ExStVCx
Gaze Locked
Gaze Dep
Form
Form/
Shape
Auditory
Broad
F
F
So far: ...
•The digital neuron model: an
adaptive mapping device.
•The neural module: capable of
depiction and memory
•The neural architecture: capable
of depictive awareness
Digital Models of the neural
brain provide ...
.. an understanding of the
mechanisms that are ESSENTIAL
for any organism to be conscious.
There appear to be FIVE
1
Perceptual mechanisms that
Depict the world
2
Imaginational mechanisms
that recall Depictions of the
world
Note: 1 and 2 mingle to
provide a sensation of the
world.
Aleksander, I. & Dunmall, B. (2000) Proc R Soc Lond B, 267, 197-200
3
Attentional mechanisms that
select which parts of the world
are Depicted
4
Planning mechanisms that cause
imaginational depictions to
predict events.
5
Affective mechanisms that
evaluate planning depictions.
These axioms lead to a
depictive architecture that
may be necessary in objects
that could be said to be
conscious.
These mechanisms are
realisable by the
neurocomputation we have
seen.
They are also the TEST for
the presence of a minimal
form of consciousness
A take-home challenge:
When the five mechanisms are
instilled in some robot,
what reason would there be for
denying that the robot is
conscious?
The issue for today
How does
brain modelling
differ
Useful
properties
emerge
from
AI programs
and neural
afrom
specific
architecture,
these
includenetworks?
awareness and
imagination
Gets
closer
to
human-like
What can it achieve?
performance in computing
systems and robots.
Contacts & Background:
email: [email protected]
NRM: [email protected]
Books:
Aleksander/Morton: Introduction to Neural
Computing (2nd ed.), Thompson Press, 1995
Aleksander: Impossible Minds: My Neurons My
Consciousness, ICPress, 1996
Aleksander: How to Build a Mind: Machines
with Imagination: Orion Press, 2001