Neural Networks, 1997
A Neural Global Workspace Model for
Conscious Attention
James Newman, Bernard J. Baars and Sung-Bae Cho
Soft Computing Laboratory
김희택
Artificial Consciousness
• Neuroscience
– ‘의식이란 두뇌의 여러 부분들의 상호작용을 통해 생성되는 상승효과’
– Neuronal correlates of consciousness (NCC)
• Monism and Dualism
– Monism
• 인간은 두뇌에서 직접 사물을 지각하고 정보를 처리
– Dualism
• 인간이 지각하는 것은 일종의 정신
• 두뇌는 정보처리에 얻은 데이터 보관장소
• Attention
– 어떤 것을 선택하여 집중하고 다른 것들을 무시하는 인지과정
– Unconscious attention 과 conscious attention
1
Global Workspace (GW)
• Simple cognitive architecture in which many parallel experts
interact via a serial
– Brain viewed as a massive parallel distributed system of highly
specialized processors (Brain web)
– Control may take place by way of a central information exchange
– Consciousness is the primary agent of such a global access function
in humans
• A theater metaphor
– Spotlight : Consciousness
– The rest of the theater : unconscious
– Control the bright spot : selective attention
• This paper present a neurocognitive model based on GW
2
Modeling Global Competitive Attention
• Consciousness
– Dynamic process, not a static structure
– Not localized to some ‘brain center’ but arises at widely distributed
networks of neurons
– Consciousness is achieved by allocating resource to specific brain
portion
• Involves a higher order of resource allocation
• The allocation of resources in case of non-conscious attention
– Visual cortex process
• Contour, movement, color, spatial location
• But human’s awareness is of a single, coherent object
– ’binding problem’ of how these multifarious representations are
integrated into real-time ‘objects’ of perception
3
Neural Mechanism for Global Attention
• Parallel loops of thalamo–cortical–thalamic axons, passing
through a thin sheet of neurons known as the nucleus
reticularis thalami (nRt)
• The neurons of nRt are largely GABA-ergic, inhibitory neurons
• There is an orderly topography to this array of axon collaterals
and underlying thalamic nuclei
– It essentially mirrors, in miniature, the modular architecture of the
cortex
Nucleus reticularis thalami (시상망상핵)
이는 간뇌피개와 중뇌피개의 망상체핵의 계속이며, 시상수
판 내에 있는 수판내핵(Nuclei intralaminares)도 이와 같은
성질의 것이다. 흔히 내측핵에 속하던 Nucleus median, 또
는 Nucleus centrum medianum은 P.N.A에서는 중심내측핵
(Nucleus medialis centralis)이라 하나, 그 기능은 망상핵군
과 같이 취급된다.
4
Main model of the nRt
• Modeling a competitive
network for global attention
based upon a highly
Thalamus (시상)
simplified version of the
뇌의 제3뇌실 측벽의 대부분을 이루는
1쌍의 커다란 알 모
‘thalamus–nRt–cortex
양의 기관. 뇌의 5개 부분 중 하나인
간뇌(間腦)의 주요구조
complex’
이다. 시상은 뇌간(腦幹)과 전뇌 사이에 위치하고, 고삐
(habenula)와 송과선을 포함하는 시상상부(epitha-lamus),
• Takes into와account
자율신경과 연결된 시상하부(hypothalamus)
연결되어the
있다. 시상에는 2가지 중요한 기능이
있다.of
시상은
각종
effects
dendro–dendritic
수용기로부터 신경충격을 대뇌피질로
전달하는데,
그곳에
interactions
throughout
nRt
서 신경충격은 깨어 있는 동안 촉각·통증·온도 등의 적절한
– Provided the basis for a
감각으로 경험된다. 또한 시상은 휴지상태중에는 시냅스 전
simple version of the global
달, 즉 들어오는 충격을 조절하는데, 이러한 조절은 긴 휴지
gating model
기 이후에 오는 폭발적인 신경충격을 분산시킴으로써 이루
진다.
5
Simulation Result of nRt model
• The overall pattern of activation is exclusively dependent upon
the wave pattern of the NRT units
– nRt is the source of global oscillatory activity
• The synchronous activation of specific and non-specific loops is
postulated to provide a basis for ‘perceptual unity’ by which
different sensory components are gathered into one global
image
6
A Neural Model for Global Resource Allocation
Cerebrum cortex
• Wagon wheel model
– Contributing to global
attention an conscious
perception
nRt
Thalamus
• Two key aspects of this
‘conscious system’
– Polymodal, integrating
sensory inputs, but motor
and ‘higher-order’
cortical effects
– Extended, with
input/output relations
reaching from the brain
stem core to association
cortices
Sensory System
7
Global Resource Allocation Method
• Global attention is acting upon a highly selective stream of
information
• Executive system is that it acts upon other sub-systems,
modifying their inputs for its particular purposes
• Present a simple, but highly relevant, connectionist model that
simulates a GW system employing a gating network to
selectively filter and integrate inputs as a function of central
knowledge stores
8
Programmable Blackboard Model for Reading
Central
word
Central
letter
Programmable
letter
Programmable
word
9
Combine the Result of the Individual Network
• The expert networks compete each other
– The expert networks may be trained separately
• The gating network mediate this competition
• While a single expert execute the task, an automatic (‘nonconscious’) processing are operated in instances
10
Conclusion & Discussion
• Introduced neuroscience and NN models for attention, binding,
and resource allocation
• Share important features and parallels with a Neural Global
Workspace System for conscious attention
• 기계가 표면적으로는 사람과 같은 의식을 가진 행동을 보이더라도,
그것으로 정말로 ‘기계가 정신을 가졌다’ 라고 말할 수 있는가?
• 대뇌의 물리적 구조를 모델링 하는 접근방식으로 의식을 가진 기계를
만들 수 있을까?
• Conscious Attention 을 실제 어떤 분야에 응용할 수 있을까?
11
E.N.D