An Artificial Mind via Cognitive Modular Neural Architectur

1.

An Artificial Mind
via
Cognitive Modular Neural Architectur
Pentti O. A. Haikonen
NOKIA RESEARCH CENTER
P.O. Box 407
FIN-00045 NOKIA GROUP, Finland
[email protected]
© Pentti O A Haikonen / Nokia Research Center 2000

2.

What is Mind?
Human mind is characterized by
Inner imagery,
Inner speech,
Sensations,
Emotions
An artificial mind should also have these
to qualify as a mind of any credibility.
© Pentti O A Haikonen / Nokia Research Center 2000

3.

Human Mind
-The human brain processes information with meaning
and importance
-There is “an unified experience” the instantaneous
sensory information from multiple sensors is bound
together and is coupled to the system’s knowledge and
emotional state so that a stream of interpretation,
meaning and mental responses arises - the flow of
inner imagery, inner speech, feelings
-This style of information processing -cognitive
information processing- is completely different from
present day computers
© Pentti O A Haikonen / Nokia Research Center 2000

4.

What is Involved in Cognition?
-Meaning and understanding
-Perception and recognition
-Prediction
-Priming
-Attention
-Match/mismatch/novelty detection
-Learning and memory
-Judgement, good/bad
-Pain and pleasure
-Emotions
-Motivation, needs, drives, goals
-Deduction, reasoning, planning
-Language
-Consciousness?
© Pentti O A Haikonen / Nokia Research Center 2000

5.

What is Involved in Consciousness?
-Awareness of environment
-Awareness of own body
-Awareness of qualia, how it feels
-Introspection of thoughts, emotions and feelings
-Awareness of past, present and expected future
-Awareness of self, one’s own existence
-Awareness and ability to report the existence of
one’s inner imagery and speech as such
© Pentti O A Haikonen / Nokia Research Center 2000

6.

Inner Imagery, Inner Speech
-The flow of inner speech, inner imagery is typical to
human cognition
- Inner speech and inner imagery are also understood as
verbal and visual thinking
-Inner speech, imagery, emotions and sensations are also
the contents of our consciousness
© Pentti O A Haikonen / Nokia Research Center 2000

7.

Steps towards Machine Mind
1. Devise suitable information representation
method (distributed signal representation)
2. Devise an elementary processing unit for the
above (non-numeric associative neuron)
3. Devise system architecture that can support inner
imagery etc. and the cognitive processes
(reentrant modular architecture)
© Pentti O A Haikonen / Nokia Research Center 2000

8.

The Neuron
input signal
threshold
output signal
match signal
associative signal array
m signals
-
mismatch signal
novelty signal
Preservation of the input signal meaning
Correlative Hebbian (Associative) learning
Resolves match/mismatch/novelty states
Non-numeric
© Pentti O A Haikonen / Nokia Research Center 2000

9.

Neuron Group as the Basic Signal
Processor
n signals
NEURON GROUP
THRESHOLD
input signal array
output signal array
n signals
match signal
associative signal array
mismatch signal
m signals
novelty signal
- Association of signal arrays to each other
- Associative evocation of output signal arrays
- Compression or generalization when n < m
- Amplification or priming by the associative signal array
- Resolves match/mismatch/novelty between input and evocation
© Pentti O A Haikonen / Nokia Research Center 2000

10.

The Reentrant Loop -Key to Inner Imagery
sensed
representation
feedback
neurons
neuron
input
representation
evoked
representation
percept
sensor
&
neuron
object feature
detectors (feature signals)
percept
neuron
percept
associative
neuron
system
output
-Perception with and without priming
-Reverberating short term memory
-Translation of output representations into percepts
(inner imagery, inner speech), Introspection
-Grounding of meaning
-Percept - the “official” output to other modules
© Pentti O A Haikonen / Nokia Research Center 2000

11.

The Cognitive System
feedback
neurons
preprocess
percept
association neuron groups
m/mm
m/mm/n
reward
P/DP
system
punish
Sensory module 2
sensory
input
pleasure/
displeasure
system
preprocess
m/mm
system
M/MM
system
m/mm
m/mm/n
feedback
neurons
Associative
percept
association neuron groups
WTA
sensory
input
WTA
Sensory module 1
The complete system consists of:
-Multiple associatively cross-connected sensory modules
-Pleasure/displeasure system
-Match/mismatch/novelty detection
© Pentti O A Haikonen / Nokia Research Center 2000

12.

text
input
preprocess
feedback
neurons
pleasure/
displeasure
system
punish
association neuron groups
word
m/mm/n
reward
percept
WTA
Linguistic system
m/mm
m/mm
system
Associative
M/MM
system
P/DP system
shape
feedback
neurons
size
feedback
neurons
color
Visual
attention
focus
system
gaze
direction
pos
percept
association neuron groups
shape
percept
association neuron groups
WTA
feedback
neurons
association neuron groups
WTA
attention &
preprocess
association neuron groups
WTA
visual
input
m/mm
m/mm/n
WTA
Visual system
size
percept
color
percept
pos
© Pentti O A Haikonen / Nokia Research Center 2000

13.

The Simulation System
© Pentti O A Haikonen / Nokia Research Center 2000

14.

o w o rd (1 5 6 )
te x t
in p u t
feedback
neurons
D
D
D
p re d w (1 5 6 ,4 6 8 )
word(156)
m /m m /n
p rim w 1 (4 6 8 )
m
w n w 1 (1 5 6 ,4 0 )
x
m /m m
sha p e
pl easure/
p u n is h di spl easure
system
p rim w 2 (4 6 8 )
m
w n w 2 (1 5 6 ,3 )
x
m /m m
s iz e
m /m m /n
p rim w 3 (4 6 8 )
m
x
w n w 3 (1 5 6 ,3 )
c o lo r
m /m m
p rim w 4 (4 6 8 )
m
w n w 4 (1 5 6 ,4 6 )
x
m /m m
name
m
w n w 5 (1 5 6 ,5 )
pos
m /m m
m
w n w 6 (1 5 6 ,2 )
p /d p
m /m m
M /M M
s y s te m
m /m m /n
Winner-Takes-All
re w a rd
p rim w 5 (4 6 8 )
x
p rim w 6 (4 6 8 )
x
p rim w 7 (4 6 8 )
m
w n w 7 (1 5 6 ,2 )
M /M M
x
p/shape
feedback
neurons
fe a tr(4 0 )
shape
ftrp w (4 0 ,1 )
dp/shape
WTA
ftrd p w (4 0 ,1 )
m /m m
ftrn w 1 (4 0 ,1 5 6 )
word/shape
m /m m
ftrn w 2 (4 0 ,3 6 )
focus pos/shape
m /m m /n
s iz (3 )
si ze
m /m m /n
s n w (3 ,1 5 6 )
word/si ze
feedback
neurons
c o lr(3 )
col or
c n w (3 ,1 5 6 )
word/col or
p o s (3 6 )
focus pos
fp w (3 6 ,4 0 )
focus pos
WTA
m /m m
feedback
neurons
WTA
m /m m
change det.
w e n w (2 ,1 5 6 )
stepper
WTA
vi sual i nput
&
preprocess
© Pentti O A Haikonen / Nokia Research Center 2000

15.

The Simulation System
Naming entities
1. Point the
object with a
laser pointer
2. Type in a
name
3. Push
“Emph” and
“Enter”
time
© Pentti O A Haikonen / Nokia Research Center 2000

16.

The Simulation System
Teaching categories; category “shape”
© Pentti O A Haikonen / Nokia Research Center 2000

17.

The Simulation System
Naming an entity with shape, color and size attributes
© Pentti O A Haikonen / Nokia Research Center 2000

18.

The Simulation System
Deduction by evoked inner imagery, answering a question
The word “square” has not been explicitly associated to “dollar”!
© Pentti O A Haikonen / Nokia Research Center 2000

19.

The Simulation System
Deduction by evoked inner imagery, answering a question
The word “green” has not been explicitly associated to “dollar”!
© Pentti O A Haikonen / Nokia Research Center 2000

20.

The Simulation System
Deduction by evoked inner imagery, contradiction detection
© Pentti O A Haikonen / Nokia Research Center 2000

21.

The Simulation System
Deduction by evoked inner imagery, affirmation detection
© Pentti O A Haikonen / Nokia Research Center 2000

22.

The Simulation System
Emotional significance;
detection of an emotionally significant entity from noise
© Pentti O A Haikonen / Nokia Research Center 2000

23.

The Simulation System
Visual search of a given entity; the search is completed when
a sensed object matches the inner image of the object to be
searched. No pattern matching is done however!
© Pentti O A Haikonen / Nokia Research Center 2000

24.

The Simulation System
-Verbal sequences, reproduction
-Sequences as serial associative prediction; detection of
mismatch between prediction and actual percept
© Pentti O A Haikonen / Nokia Research Center 2000

25.

Conclusions 1
A modular non-numeric neural network has been
devised that
-Operates with inner imagery, inner speech
-Acquires information via perception
-Acquires information about its inner states via
introspective perception
-Is able to learn and generalize (and fast!)
-Has cognitive functions similar to human brain
What remains to be demonstrated:
-Actual motor output systems
-The effect of needs, drives, planning, will
-Personal history, sense of time
-Self concepts
© Pentti O A Haikonen / Nokia Research Center 2000

26.

Conclusions 2
TOWARDS CONSCIOUS MACHINES?
This system has the flow of inner imagery and inner
speech, the hallmarks of human consciousness
However, the system is not yet able to report on its
own that it exists, that it has inner imagery and inner
speech
A system’s ability to report on its own that it has inner
speech, produced by the system self, could be used as
a test for machine self-consciousness
The author would like to see the Turing test be
replaced by this one.
© Pentti O A Haikonen / Nokia Research Center 2000

27.

Thank You !
An Artificial Mind via
Cognitive Modular Neural Architecture
text
input
preprocess
feedback
neurons
reward
percept
pleasure/
displeasure
system
punish
association neuron groups
word
m/mm/n
WTA
Linguistic system
m/mm
m/mm
system
Associative
M/MM
system
P/DP system
feedback
neurons
size
feedback
neurons
color
gaze
direction
pos
association neuron groups
WTA
percept
shape
percept
association neuron groups
WTA
feedback
neurons
shape
Visual
attention
focus
system
m/mm
m/mm/n
attention &
preprocess
NOKIA RESEARCH CENTER
size
percept
association neuron groups
WTA
visual
input
color
percept
pos
P.O. Box 407
FIN-00045 NOKIA GROUP Finland
association neuron groups
WTA
Visual system
Dr. Pentti O. A. Haikonen
principal scientist
Cognitive Technology
[email protected]
© Pentti O A Haikonen / Nokia Research Center 2000