Transcript Document

Progress in Machine consciousness
David Gamez, Consciousness and Cognition
vol.17, pp.887-910, 2007
Software Agent
2009. 04. 09
Seunghyun Lee
Contents
• Introduction
• Classification of Machine Consciousness
– MC1, MC2, MC3, and MC4
• Research Projects
• Relationship with Other Areas
• Criticisms
• Issues and Potential Benefits
1
S
FT COMPUTING @ YONSEI UNIV . KOREA
Introduction
• Machine consciousness
– Test theories of consciousness using computer models
– Create more intelligent machines that might actually have
phenomenal states
– “Artificial consciousness”, “digital sentience”
• Breaking machine consciousness into four areas
Category
Associated Subject
MC1
External behavior
MC2
Cognitive characteristics
MC3
Architecture
MC4
Phenomenally consciousness
2
S
FT COMPUTING @ YONSEI UNIV . KOREA
Classification of Machine Consciousness
• MC1(External behavior)
– Goal
 Replicate conscious human behavior
Unconscious behavior
Conscious behavior
Feature
-Automatically carried out
-Limited amount of behavior
-Complex activities
-New behaviors can only be learnt
when consciousness is present
Example
Muscle contractions while walking,
epileptic seizure
Driving home from work,
interpersonal dialog
– Large lookup table / zombie robot
– Example
• Turing Test
3
S
FT COMPUTING @ YONSEI UNIV . KOREA
Classification of Machine Consciousness
• MC2(Cognitive characteristics)
– Goal
 Research on connection between consciousness and cognitive
characteristics
– Imagination, emotion, and self
– Metzinger’s 11 constraints on consciousness
(1) Global availability
(2) Activation within a window of presence
(3) Integration into a coherent global state
(4) Convolved holism
(5) Dynamicity
(6) Perspectivalness
(7) Transparency
(8) Offline activation
(9) Representation of intensities
(10) “Ultrasmoothness”: Homogeneity of simple content
(11) Adaptivity
– Alexsander’s five cognitive mechanisms
4
S
FT COMPUTING @ YONSEI UNIV . KOREA
Classification of Machine Consciousness
• MC3 (Architecture)
– Goals
 Simulation of architectures related to human consciousness
– Global workspace(Baar), neural synchronization(Crock)
• MC4(Phenomenally consciousness)
– Phenomenally consciousness?
• We have phenomenally conscious states when we see, hear, smell, taste and have p
ains. (Block 1995: 230)
– Goals
Research on machines that have real phenomenal experiences
that are actually conscious themselves
– System based on biological neurons
5
S
FT COMPUTING @ YONSEI UNIV . KOREA
Axioms and Neural Representation Modeling
Research Project
• Five axioms(Alekxsander, Dunmall, 2003)
Depiction
Represent elements of the world(perceptual states)
Imagination Recall parts of the world or create sensations
Attention
Select which parts to be depicted or imagined
Planning
Control sequences of states to plan actions
Emotion
Evaluate planned actions and determine the ensuing action
• Kernel Architecture(Alekxsander, 2005)
6
S
FT COMPUTING @ YONSEI UNIV . KOREA
CRONOS
Research Projects
<CRONOS>
<SIMNOS>
• Main focus of this project
– Cognitive, architectural and phenomenal aspects of machine
consciousness (MC2~4).
• Constitution
– CRONOS, SIMNOS, biologically inspired visual system, SpikeStream
7
S
FT COMPUTING @ YONSEI UNIV . KOREA
CRONOS
Research Projects
• Approach 1(Holland, 2003)
– Focus on internal model
– Test
•
•
•
•
SIMNOS as an internal model of CRONOS
CRONOS eyes obtains information from environment
Update SIMNOS
Internal model : ‘offline’  ‘imagine’ mode before selected action is carried out
• Approach 2(Gamez)
– Development of spiking neural network that controls eye movement
• Generates eye movement spontaneously to the different part
• Learns the association between the eye’s position and a visual stimulus
– Emotional system
• Negative object  ‘imagination’ mode  inhibit sensory input and motor output
– Cognitive characteristic(MC2), neural correlated architecture(MC3)
8
S
FT COMPUTING @ YONSEI UNIV . KOREA
Cog
Research Projects
• Brooks, Breazeal et al.(1998)
• Constitution
– 4 cameras, 2 microphones, and
many piezoelectric touch sensors
– A number of hard wired innate reflexes
– Emotional System
• Independent projects
– Joint attention, theory of mind, social
interaction, dynamic human-like arm motion,
and multi-modal coordination
• Relation
– Joint attention, theory of mind (MC1)
– Cog’s emotional system(MC2)
• Limit
– Many individual human behaviors are implemented
– Active all together  incoherence and interference
S
FT COMPUTING @ YONSEI UNIV . KOREA
9
CyberChild
Research Projects
• Simulated infant(Cotterill, 2000)
• Controlled by a biologically inspired
neural system
– Premotor cortex, supplementary motor
cortex, frontal eye fields, thalamic nuclei,
hippocampus and amygdala
– Interconnection : based on anatomical
connectivity
• Simulation
– Blood glucose measurement, milk, urine
– Sustain avoiding discomfort
• Goal
– Identify neural correlates of consciousness
• Relation
– Neural correlates of consciousness(MC3), (MC4)
10
S
FT COMPUTING @ YONSEI UNIV . KOREA
Khepera models
Research Projects
• Approach 1(Holland and Goodman, 2003)
– Test the role of internal model in consciousness
– Using ARAVQ(Adaptive Resource-Allocating Vector Quantizer)
– Graphical representations of inner states
– Experiment
• Wall following and obstacle avoidance behavior
• ARAVQ build up concepts  forms internal model
• Good performance
– Some of the internal models in humans are integrated into
conscious cognitive states(MC2)
11
S
FT COMPUTING @ YONSEI UNIV . KOREA
Khepera models
Research Projects
• Approach 2(Ziemke et al., 2005)
– Imagination
– Using simple neural network
– Constitution
• Sensorimotor module : Avoid obstacle, perform fast straightforward motion
• Prediction module : Predict the sensory input of the next time step
– ‘Imagined’ sensory inputs produced very similar behavior to real
sensory input
– MC2
12
S
FT COMPUTING @ YONSEI UNIV . KOREA
Global Workspace Models
Research Projects
• Global workspace theory(Baar, 1988)
13
S
FT COMPUTING @ YONSEI UNIV . KOREA
Global Workspace Models
Research Projects
• IDA naval dispatching System(Franklin, 2003)
– Assign sailors to new billets
– Functions are carried out using codelets
– Apparatus for consciousness
•
•
•
•
Coalition manager
Spotlight controller
Broadcast manager
A number of attention codelets
– MC1, MC2, MC3
• Neural simulations(Dehaene et al., 1998)
– Stroop task
• Predictions about brain imaging patterns
– Attentional blink
• Explained using the theory about the implementation of a global workspace in the
brain
– MC2, MC3
14
S
FT COMPUTING @ YONSEI UNIV . KOREA
Global Workspace Models
Research Projects
• Brain-inspired cognitive architecture(Shanahan, 2006)
– Functionally analogous components
to brain structure
– Enable the system to follow
chains of association
– Explore the potential
consequences prior to the action
– Experiment
• Webot and Khepera robot
• Low level actions and preferences for cylinders with different color
– Produced behavior(MC1), imagination and emotion(MC2), based on
global workspace model(MC3)
15
S
FT COMPUTING @ YONSEI UNIV . KOREA
Language and Agency
Research Projects
• Language and agent-based architecture(Angel, 1989)
– Three attributes for conscious system
1. Independent purpose regardless of its contact with other agents.
2. The ability to make interagency attributions on a pure or natural basis.
3. The ability to learn from scratch significant portions of some natural language, and
the ability to use these elements in satisfying its purposes and those of its
interlocutors.
– Nobody has implemented with this model
• Inner speech(Steels, 2003)
– Experiments in which two robotic heads watched scenes and played
a language-game that evolved a lexicon or grammar
– Rehearse future dialogue, submit thoughts to self-criticism, and
conceptualize and reaffirm memories of past experiences
– MC2
16
S
FT COMPUTING @ YONSEI UNIV . KOREA
Cognitive Architecture
Research Projects
• Cognitive approach(Haikonen, 2003)
– System intended to develop
emotion, transparency, imagination,
and inner speech
– Sensory modules
– Main idea
Percepts  Conscious
different modules cooperate in
unison, focus on the same entity,
forms associative memories
– MC1~4
17
S
FT COMPUTING @ YONSEI UNIV . KOREA
Cognitive Architecture
Research Projects
• Schema-based model(Samsonovich and DeJong, 2005)
– Based around schemas
– Constrained by a set of axioms
– Axioms correspond to the
system’s ‘conscious’ self
– MC1, MC2, but not MC3
• Cicerobot(Chella and Macaluso)
– Museum tour guide robot
– Based around an internal 3D simulation
 plan actions
– Conscious cognitive architecture(MC2),
control the robot(MC1)
18
S
FT COMPUTING @ YONSEI UNIV . KOREA
Research Projects
• Synthetic phenomenology
– New area of research on machine consciousness
– Develop artificial systems which are capable of conscious states and
the description of their phenomenology when and if this occurs
– Challenges
• Develop systems which be capable of phenomenal states
“To be synthetically phenomenological, a system S must contain machinery that
represents what the world and the system S within it seem like, from the point of
view of S’’(Aleksander and Morton, 2006)
• Find ways of describing phenomenal states when and if they
occur
Graphical representations of Kheperas’ inner states(Holland, 2003)
• Distinguish machine’s phenomenal and non-phenomenal states
which internal states are likely to be conscious?
19
S
FT COMPUTING @ YONSEI UNIV . KOREA
Relationship with Other Areas
• Strong and weak AI(Searle, 1980)
– Weak AI : Powerful tool when we study mind (modeling)  MC1~3
– Strong AI : Programmed computer is mind itself  MC4
• Artificial general intelligence
– Goal : Replicate human intelligence completely
• ex) chess playing
– MC1 : Conscious human behavior
• Psychology, neuroscience and philosophy
– Psychology : Build also computer cognition model not only
conscious state but also others
– Neuroscience : Trend that tests theories about attention and
consciousness with neurons
– Philosophy : Common in the use of logic
S
•
FT COMPUTING @ YONSEI UNIV . KOREA
20
Criticisms
• The hard problem of consciousness
– Easy problem : Discriminate, integrate information, report mental
states, focus attention etc…(MC1, MC2, MC3)
– Hard problem : Explaining phenomenal experience(MC4)
 Many theories, but no real idea to solve
• Consciousness is non-algorithmic
– Processing of an algorithm is not enough to evoke phenomenal
awareness because of subtle and largely unknown physical
principles
• What computers still cannot do
– Fact based system cannot solve human intelligence which depends
on skills, a body, emotions, imagination and other attributes that
cannot be encoded into long lists of facts
21
S
FT COMPUTING @ YONSEI UNIV . KOREA
Potential Benefits and Issue
• Potential benefits
– MC1 : Help people to produce more imitation of human behavior
ex) chatterbots
– MC2 : Machine which understand human world and language in a
human-like way can assist people
– MC3 : Help people to understand how the brain processes
information, so that it is able to develop prosthetic interfaces to
restore visual, auditory or limb functions
– MC4 : Help people to understand the phenomenal states of very
young or brain-damaged people who are incapable of
communicating their experiences in language
22
S
FT COMPUTING @ YONSEI UNIV . KOREA
Potential Benefits and Issue
• Issues
– Can machines take over and enslave humans?
– How we should treat conscious machines?
– How should it be the legal status of conscious machines?
• Discussion
23
S
FT COMPUTING @ YONSEI UNIV . KOREA