Abstract for AISB 2000: How to Design a Functioning Mind

Abstract for the
Symposium on How to Design a Functioning Mind
17-18th April 2000
At the AISB'00 Convention


AUTHOR: Pentti O. A. Haikonen
Affiliation: Nokia Research Center

TITLE: "An Artificial Mind via Cognitive Modular Neural Architecture"

ABSTRACT:

The author proposes that an artificial mind should be able
to duplicate the processes of the human mind, i.e. inner
imagery, inner speech, sensations, the cognitive functions
like introspection, perception, attention, match, mismatch
and novelty detection, learning, memory, reasoning,
planning, emotions and motivation, perhaps even
consciousness.

An artificial PC-based cognitive system with real digital
camera and text input based on these requirements is
presented here. This neural network-based system utilizes
distributed signal representation, sensory preprocessing
into feature signals, processing of information
associatively with meaning and significance and a modular
reentrant architecture that allows the establishment of
inner imagery and speech as well as introspective perception
of these. Since the information is not represented by
numbers or number arrays no numeric computing takes place.
No pattern matching algorithms are used either. Match,
mismatch and novelty signals are derived from neuron-level
signal relations and they are used to effect sensory and
inner attention. Pleasure/displeasure conditions are also
modeled and they contribute to the reactive state of the
system.

The modular system architecture consists of
perception/response reentrant loop modules for linguistic,
visual and gaze direction subsystems as well as subsystem
modules for pleasure/displeasure and match/mismatch
evaluation. The perception/response reentrant loop realizes
sensory perception primed by prediction and inner
evocations, introspective perception and the establishment
and the grounding of meaning of inner imagery and inner
words. Additionally the reentrant loop acts as a
reverberating short-term working memory. In-loop associative
neuron groups facilitate associative crossconnections to
other modules. Learning and long-term memory are realized
via synaptic strength modifications. Each reentrant
perception/response loop operates on its own, but other
modules may evoke representations in other modules via the
associative crossconnections. However, these representations
have to compete against the others and only the strongest
ones will be eventually perceived by the target modules.

The system can learn to recognize figures, learn the meaning
of concrete words by ostension and via correlation, learn
certain abstract words and rudimentary syntax by examples,
learn to recognize new figures by verbal description, learn
temporal sequences and predict their continuation, detect
affirmation and contradiction, deduct the properties of a
given object from evoked inner imagery. Learning by
imitation would be possible if real motor output units were
given. Learning is inductive and fast, only few repetitions
are needed.

This system has several features that are commonly
attributed to consciousness: It is perceptive, the meaning
of the signals is grounded to sensory information; it has
inner imagery and inner speech; it is introspective, the
inner workings are perceived by the system via reentry to
perception process; there is distinction between the
externally evoked and internally evoked inner
representations; there is cross-module reportability, the
activity of one perception/response reentrant loop module
can be reported by one or all the other modules in their own
terms; there is attention and short-term memory. However, at
this moment the experimental cognitive system does neither
have a body reference for self-concept nor episodic memory
capacity for personal history, these will have to be added
later.

===========================================================

Pentti O A Haikonen is principal scientist in cognitive
technology at Nokia Research Center, Helsinki, Finland.
Since 1994 he has studied artificial cognition with computer
simulations. He has also long term experience on digital
image processing and compression as well as AF, RF and video
circuits design, both analog and digital. At this moment he
has 12 granted patents.

Some recent artificial cognition related publications and
papers:

Haikonen Pentti O. A. (1999), "An Artificial Cognitive
Neural System Based on a Novel Neuron Structure and a
Reentrant Modular Architecture with Implications to Machine
Consciousness", Dissertation for the degree of Doctor of
Technology, Helsinki University of Technology Applied
Electronics Laboratory, Series B: Research Reports B4, 1999

Haikonen Pentti O. A. (1998), "An Associative Neural Model
for a Cognitive System" in Proceedings of International
ICSC/IFAC Symposium on Neural Computation NC'98 Vienna 23 -
25 September 1998 pp. 983 - 988

Haikonen Pentti O. A. (1998), "Machine Cognition via
Associative Neural Networks" in Proceedings of Engineering
Applications of Neural Networks EANN'98 Gibraltar 10 - 12
June 1998 pp. 350 - 357

Haikonen Pentti O. A. (1994), "A Novel Neuron for
Associative Neural Networks", in Conference on Artificial
Intelligence Research in Finland STep-94, Proceedings of
Contributed Session Papers pp. 9 - 12, Suomen Tekošlyseura
ry, Helsinki

Haikonen Pentti O. A. (1994), "Towards Associative
Non-algorithmic Neural Networks" in Proceedings of IEEE
International Conference on Neural Networks ICNN'94 June 28
- July 2, 1994 Vol II pp. 746 - 750