Transcript From Machine Cognition to Conscious Machines

Exystence Workshop
Machine Consciousness: Complexity Aspects
From Machine Cognition to
Conscious Machines
Dr. Pentti O A Haikonen,
Principal Scientist, Cognitive Technology
Nokia Research Center
© NOKIA
exymach / 10.9.2003 / PHa page: 1
The Outline of the Story
Why machine cognition?
About intelligence, understanding and cognition
A model for cognitive and thinking machines
Machine cognition and consciousness;
coins or egg and chicken?
© NOKIA
exymach / 10.9.2003 / PHa page: 2
-one sided
Towards More Sophisticated Information
Technology Products
Many of today’s products are so complicated that the user
cannot possibly control all the involved processes, the product
itself must take care of these and leave only higher level
control and decisions to the user.
This is achieved by preprogrammed rules and algorithms that
are sometimes called embedded intelligence.
It is seen that blind algorithms are only a limited solution, more
complex context awareness, machine understanding, is
required.
© NOKIA
exymach / 10.9.2003 / PHa page: 3
Levels of Machines; One more Step to Go
The machine as a tool:
-The user can execute an action with the machine
The machine as an automaton:
-The user (or condition) can initiate action sequences
The machine as an agent:
-The user can request a context dependent function
The machine as an autonomous agent:
-The machine executes context dependent actions
as is deemed necessary by (heuristic) set of rules
The machine as a cognitive agent:
-The machine understands and is aware; is able to
execute tasks requiring real intelligence and thought
© NOKIA
exymach / 10.9.2003 / PHa page: 4
Machine Understanding of Meaning and
Context Will Enable New Possibilities
Speech and language; recognition, understanding and translation
Vision; visual episode understanding, prediction
Unified sensory information understanding
Non-indexed data base search and information compilation
Improved Artificial Intelligence, artificial creativity
Personal artificial assistants and companions
Autonomous robots;
sterile nurses that do not fall ill, rescue robots, etc.
Security, defense and law enforcement
Etc.
© NOKIA
exymach / 10.9.2003 / PHa page: 5
What We Want:
Cognitive Information Technology
or
Human-like Information Processing
From speech recognition to speech understanding
From pattern recognition to scene understanding
From text parsing to story understanding
From statistical "learning" to cognitive learning
From numerical simulation to free imagination
© NOKIA
exymach / 10.9.2003 / PHa page: 6
Traditional Models and Tools
GOFAI - The brain is a computer; human-like intelligence
and cognition via programs and algorithms?
Artificial Neural Networks; human-like intelligence via
statistical computing?
DSP; Systems with sensors, sensory information
represented by numeric values - processing by transforms,
filtering, etc. with partly parallel architectures.
Semantic Networks; understanding via classification and
indexing?
© NOKIA
exymach / 10.9.2003 / PHa page: 7
Traditional Methods have not Provided
True Intelligence and Understanding
Humans surpass the computer in everyday tasks
because humans are intelligent and are able to
understand.
But, what is intelligence and understanding?
© NOKIA
exymach / 10.9.2003 / PHa page: 8
What is Intelligence?
Instruction
booklet
No intelligence is needed if you can use this by
following the instructions only
© NOKIA
exymach / 10.9.2003 / PHa page: 9
What is Intelligence
Intelligence is what we use when rules do not help.
This excludes the possibility of rule-based AI
© NOKIA
exymach / 10.9.2003 / PHa page: 10
Understanding is not that Simple
For instance - Episodic Understanding:
NOT tape recorder type storage and playback
BUT the ABILITY to
-Answer questions about the subject;
-what is where
-what is happening
.
-who is doing what to whom, etc.
-Paraphrase; describe in own words
-Detect contradictions
-Predict what happens next, what is possible
-Evaluate significance, is this good or bad
-Give reasons for present situation, Etc.
© NOKIA
exymach / 10.9.2003 / PHa page: 11
Understanding Necessitates the
Grounding of Meaning
Real world concepts must be grounded to real world entities.
-This calls for a perception process
Concepts must be connected to other concepts.
-This calls for associative cross-connections
The general model of cognition provides these functions
and more.
© NOKIA
exymach / 10.9.2003 / PHa page: 12
The General Model of Cognition
Environment
objects
actions
situation
relationships
Perception process
percepts of:
-environment
-inner states
match/mismatch/
novelty detection
emotional
evaluation
Reactions
Actions
learned routines
© NOKIA
exymach / 10.9.2003 / PHa page: 13
Internal process
tasks, goals, needs
memories
experience
prediction
reasoning
planned action
judgement
emotions
etc.
The Cognitive Model as an Associative Structure
© NOKIA
exymach / 10.9.2003 / PHa page: 14
Representation of Information
or
the Power of Power Sets
The idea behind distributed signal representations:
The set of all subsets -the power set- is always larger than
the original set. Therefore a limited number of original set
members -here the basic signals- can give rise to a much
larger number of subsets -possible signal combinationsallowing thus the representation of large number of
different cases with only limited number of basic signals.
© NOKIA
exymach / 10.9.2003 / PHa page: 15
Distributed Representations and
Associative Processing
Signals and signal sets carry meaning
Individual signals correspond to elementary features
Signal sets or arrays correspond to entities
Entities can be associated together by linking the
corresponding signal arrays
An entity can be evoked by incomplete or slightly different
signal array -”the closest guess”
Also episodes, signal set sequences, can be handled
© NOKIA
exymach / 10.9.2003 / PHa page: 16
Distributed Representations and
Associative Processing Go Well Together
Provided that
Combinatorial explosion is avoided by attention; a mechanism
that limits the actual connections by relevance and
importance, etc. Hence the eventual need for emotional
significance.
Interference -the false evocation of undesired signals- is
controlled, various methods exist.
© NOKIA
exymach / 10.9.2003 / PHa page: 17
The Outline of a Conscious Machine
feedback
environment
raw
distributed
signals
sensors &
preprocessing
responses
© NOKIA
exymach / 10.9.2003 / PHa page: 18
percepts
perception
process
inner
processes
The
General
Architecture
Visual subsystem
sensor
preprocess
perception
process
percept
neuron group
perception
process
percept
neuron group
perception
process
percept
neuron group
perception
process
percept
neuron group
percept
neuron group
Auditory subsystem
sensor
preprocess
Touch subsystem
sensor
preprocess
Body position
sensor
preprocess
motor actions
threshold
System
smell, taste, pain, etc.
system reactions
© NOKIA
exymach / 10.9.2003 / PHa page: 19
perception
process
threshold
Characteristic Properties of the Model
Each modality works on its own and produces streams of
percepts about environment and internal states.
Modalities are associatively cross-connected, therefore the
activity of one modality may be reflected in the other modalities;
percepts may be named and labeled, names may evoke
corresponding percepts…the activity of one modality may be
memorized and reported in terms of other modalities, etc.
Attention determines which percepts are accepted for further
action. Attention is controlled by signal intensity and thresholds,
these are controlled by e.g. emotional significance.
Pain and pleasure are system reactions that affect attention.
The flow of inner speech and inner imagery is supported.
© NOKIA
exymach / 10.9.2003 / PHa page: 20
Consciousness in the Machine?
How do we know if the machine is conscious? There may be
some telltale symptoms that we could look for.
For instance Prof. Aleksander lists five axioms:
1) sense of place,
2) imagination,
3) directed attention,
4) planning,
5) decision/emotion
In the following a rather similar list is given perhaps with
some twists.
© NOKIA
exymach / 10.9.2003 / PHa page: 21
Consciousness in the Machine?
-Does the machine have mental content that is about something?
-Is the machine able to report its mental content to itself (and
others) and does it recognize the ownership of the same?
-Is the machine able to make the difference between the
environment and the machine self?
-Does the machine have (episodic) sense of time?
-Does the machine bind its present experience to personal history
and expected future? (“the flow of existence”)
-Is the machine aware of its own existence?
-The “hammer test” of phenomenal awareness: Does the machine
feel pain?
© NOKIA
exymach / 10.9.2003 / PHa page: 22
Consciousness in the Machine?
And finally,
if the machine were to genuinely ask:
Where did I come from?
Then we would know that we are into something deep.
© NOKIA
exymach / 10.9.2003 / PHa page: 23
Cognition and Consciousness;
Which Comes First?
In this model understanding arises from the processing with
meaning -on the other hand meaning-carrying signal arrays
are intentional, a supposed prerequisite for consciousness.
Here also the cross-modality binding and reporting -hallmarks
of consciousness- arise from the requirements of cognition.
Therefore, are true cognition and consciousness connected or
separate properties?
Thus, would the proper realization of cognition automatically
result in some kind of consciousness (Or do zombies exist)?
© NOKIA
exymach / 10.9.2003 / PHa page: 24
From Machine Cognition to Conscious Machines
Thank You
for Your Attention!
Dr. Pentti O A Haikonen, Principal Scientist, Cognitive Technology
Nokia Research Center
© NOKIA
exymach / 10.9.2003 / PHa page: 25
-end of show-
© NOKIA
exymach / 10.9.2003 / PHa page: 26