an information flow perspective on access consciousnessmozer/research...consciousness course spring...
TRANSCRIPT
An Information Flow Perspective on Access Consciousness
Michael C. MozerDonald W. MathisJennifer Finger
Michael ColagrossoInstitute of Cognitive Science andDepartment of Computer ScienceUniversity of Colorado, Boulder
Consciousness CourseSpring 2004
Cognitive Architecture
Cortical computation can be characterized by coarse-scale, functionally specialized pathways.
E.g., visual word-form recognitionidentification of semantic features of visual objectsauditory word recognitioncomputation of spatial relationshipsconstruction of motor plans
Cognition requires coordination of multiple pathways.
Pathways act as associative memories.E.g., visual word recognition
input: visual featuresoutput: letter strings
Past experience imposes well-formedness constraintson the output.
Letter strings must be consistent with English orthography.
Output is the best-fitting interpretation of the input.
pathway
input domain
output domain
Cognitive Architecture
Pathway operation shows a speed-accuracy trade off.Initial output appears rapidly, but is inaccurate.
Pathway asymptotically converges on the best interpretation.
With each experience, pathway tends to produce its response more rapidly.
80 100 120 140 160 180 200 2200
0.2
0.4
0.6
0.8
1
processing time
probability ofcorrect output
Pathway Operation
Implementation using two standard connectionist components:
input domain output domain(e.g., semantic)(e.g., visual)
dogcat
cow rat
neutral
output
input
attractor network
feedforward network
Pathway Implementation
Speed-accuracy trade off due to• inaccuracies in feedforward mapping• Gaussian noise added to unit activities
Common connectionist architecture• Hinton & Shallice (1991); Plaut & Shallice (1993); Mozer
& Behrmann (1990)• Cascaded attractor net models incorporate feedforward
mappings (e.g., McClelland & Rumelhart, 1981).• Feedforward net models implicitly assume attractor net
for read out (e.g., Sejnowski & Rosenberg, 1987).
Pathway Dynamics
attractor unit update equation:
state unit update equation:
attractorunits state
units
externalinput
a(t)
s(t)
e(t)
: center of attractor j: width of attractor j
µj
βjaj t( ) s t( ) µj–
2 βj⁄–( )exp=
aj t( ) aj t( ) ai t( )∑⁄=
si t( ) h si t 1–( ) ωext+ ei t( ) ωattr aj t 1–( )µjij∑ ηi t σ;( )+ +=
freeparameters
Cascaded Pathways
Sequential operation?• biologically implausible
• computationally inefficient
pathway A
pathway B
classificationresponse
semantics
visual input
Information Flow
What sort of representation in the output of pathway A will support a response from pathway B?
Sufficient condition: familiar and temporally persistent representation
0.940.95
0.960.97
0.980.99
1
50
100
150
200
250
3000
0.2
0.4
0.6
0.8
1
duration of pathwayA representation
familiarity of pathwayA representation
A
Bprobabilityof correct
pathway Boutput
Information Flow
Familiar and persistent representationattained when pathway reaches attractor
Information Flow
Familiar and persistent representationattained when pathway reaches attractor
At attractor, representation is stable.
Stability defined to occur if ,where ,
,, and
d t( ) ζ<d t( ) αd t 1–( ) 1 α–( )d t( )+=d t( ) s t( ) s t 1–( )–=ζ 0.25= α 0.9=
100 120 140 160 180 200 220 240 260
100
120
140
160
180
200
220
240
260
cycles tocorrect
pathway Bresponse
cycles to pathway A stability
Level of Expertise
Stability required to initiate underlearned responses.
150 200 250 300
100
150
200
250
300
“novice”
“intermediate”
“expert”
cycles tocorrect
pathway Bresponse
cycles to pathway A stability
Task Difficulty
Stability required for tasks relying on finer granularity of information.
80 100 120 140
60
80
100
120
140
cycles tocorrect
pathway Bresponse
cycles to pathway A stability
complex
simple
(slope = .72)
(slope = –.03)
Number of Response Alternatives
Stability required for discriminations involving a large number of alternatives.
cycles tocorrect
pathway Bresponse
cycles to pathway A stability
200 alternatives(slope = .91)
50 alternatives(slope = .83)
10 alternatives(slope = .52)
2 alternatives(slope = .10)
100 150 200
50
100
150
200
250
Role of Stability
Stability of a pathway’s output is sufficient to induce the correct response from pathways to which it is connected.
Stability of a pathway’s output is necessary when
• decision making with limited domain expertise
• responses are arbitrary or complex
• responses involve discrimination among many alternatives
Stable representations are the most accessible — they allow for the flexible control of behavior (including verbal report).
A
B C D E
Neurobiological Correlate of Stability?
0.940.95
0.960.97
0.980.99
1
50
100
150
200
250
3000
0.2
0.4
0.6
0.8
1
Similar Properties for Belief Net Implementation
00.2
0.40.6
0.81
0
50
100
150
2000
0.2
0.4
0.6
0.8
1
familiarity of pathwayA representation
duration of pathwayA representation
probabilityof correct
pathway Boutput
20
40
60
80
100
120
140
160
180
200
220
70 80 90 100 110 120 130 140 150
Tim
e fo
r se
cond
pat
hway
to r
each
crit
erio
n
Time for first pathway to reach criterion
0
50
100
150
200
250
300
200 250 300 350 400 450 500
Tim
e fo
r se
cond
pat
hway
to r
each
crit
erio
n
Time for first pathway to reach criterion
120
140
160
180
200
220
240
260
280
120 130 140 150 160 170 180 190T
ime
for
seco
nd p
athw
ay to
rea
ch c
riter
ion
Time for first pathway to reach criterion
2 actions
10 actions
50 actions
Relation to Other Stories about Consciousness 1
Conscious states are interpretationsGregory: ambiguous objectsCrick & Koch: high frequency flicker not perceivedLogothetis, Lumer: rivalrous displaysMarcel: subliminal semantic priming
Attractor state is an interpretation.
Relation to Other Stories about Consciousness 2
Conscious states require high quality representations
Farah, O’Reilly, & Vecera: prosopagnosia modelMunakata
Any attractor state is a high quality representation, in the sense that adding noise moves away from attractor and also lowers quality of a representation.
Probably makes more sense to focus on familiarity than quality:Familiarity a natural way of explaining how pathway A can have an effect on pathway B.
Relation to Other Stories about Consciousness 3
Conscious states require temporal persistence
O’Brien & Opie: “stability” is a characteristic of phenomenal awarenessCrick & Koch: “coalitions” that are sustained in time; “snapshots” that must cross a threshold in timeDehaene & Naccache: “information must be maintained over a sufficient duration”
My story is about access consciousness.
Crick & Koch coalitions are probably coarser scale (front vs. back)
Relation to Other Stories about Consciousness 4
Conscious states require explicit representationsCrick & Koch, O’Brien & Opie
Relation to Other Stories about Consciousness 5
Conscious states are attractors
Rumelhart & McClelland, Smolensky: cited in Farah et al.Crick & Koch: dynamic competition among coalitions
Relation to Other Stories about Consciousness 6
Consciousness entails global availability of information to to widespread brain sources
Baars, Dehaene & Naccache: global workspace framework
My story requires functional connectivity between perceptual and response pathways for access to occur.
Frontal areas could provide this functional connectivity, so global workspace ideas complement my story.
Relation to Other Stories about Consciousness 7
Attention is a prerequisite for consciousness
Dehaene & Naccache; Crick & Koch
In my story, attention is not necessary, but it makes awareness more likely: Standard notion of attention is to amplify some representations. Attractor corresponding to amplified representation more likely to be chosen.
Semantic Priming Paradigm
Task: lexical decision to target word following prime word
Congruent conditionprime LEAF
target TREE
Unrelated conditionprime AUTO
target TREE
Facilitation on lexical decision RT for targets congruent with prime.
Modeling Semantic Priming
Presentation of stimulus strengthens attractors that are visited (McClelland & Rumelhart, 1986; Becker et al., 1997)
Nearby (similar) attractors are also strengthened.
Priming depends on state space trajectory.
Priming can occur without stability (awareness).
extendedpresentation
neutral neutral
leaf
tree
briefpresentation
neutralneutral neutral
Semantic Priming with Polysemous Words(Marcel, 1980; Swinney, 1979; Tanenhaus & Leiman, 1979)
Task: lexical decision to target
Modeling semantic primingTwo pathway architecture
Decision pathway responds “yes” if perceptualpathway obtains a well-formed state, “no”otherwise.
congruent incongruent unassociatedcontext LEAF HAND HANDprime PALM PALM RACEtarget TREE TREE TREE
decision
perceptual
orthography
semantics
lexical decision
pathway
pathway
Simulation with extended presentation of prime
After model has stabilized on PALM:• one interpretation has been selected• facilitation for targets congruent with that meaning• awareness of prime
congruent incongruent unassociatedhuman data (Marcel, 1980) 499 ms 547 ms 541 ms
simulation 224 cycles 252 cycles 257 cycles
palm1
palm2
leaf
tree
palm1
palm2
leaf
tree
palm1
palm2
leaf
tree
palm1
palm2
leaf
tree
2. present PALM1. present LEAF
wrist wrist wrist wrist
= facilitation
Simulation with brief presentation of prime
Before model has settled on PALM:• no interpretation has been selected• facilitation for targets congruent with either meaning• no awareness of prime
Human data showed correlation between onset of awareness and selection.
Model suggests further that selection is a necessary precursor to awareness.
congruent incongruent unassociatedhuman data (Marcel, 1980) 511 ms 520 ms 548 ms
simulation 238 cycles 242 cycles 254 cycles
palm1
palm2
leaf
tree
palm1
palm2
leaf
tree
palm1
palm2
leaf
tree
palm1
palm2
leaf
tree
2. present PALM1. present LEAF
wristwrist wrist wrist
Primed Stem Completion(Debner & Jacoby, 1994)
ProcedureMasked prime (e.g., TABLE)
Followed by stem to be completed (e.g., TAB_ _)
Inclusion condition: Ss instructed to complete stem with prime if possible.Exclusion condition: Ss instructed to complete stem with any word but prime.
Short (mostly unconscious) and long (mostly conscious) exposure conditions
Dependent measure: probability that stem completed with prime
Simulation
Presentation of prime (TABLE)• Perceptual pathway produces lexical representation.• Priming mechanism operates inside perceptual pathway.• If input is sustained and perceptual pathway stabilizes, lexical
representation copied to working memory.
Presentation of stem (TAB_ _ )• Stem representation: average of three similar orthographic representations• Perceptual pathway finds best-fitting completion of visual input consistent
with lexicon.• Perceptual processing modulated by contents of working memory
naming
orthography
lexical rep.
name
pathway
workingmemory
perceptualpathway
Results
Debner and Jacoby analysis of data suggests a dissociation between conscious and unconscious processes.
No such distinction in our model: awareness is an emergent property of representations, not processes or pathways.
L−I L−E S−I S−E CTL0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
L−I L−E S−I S−E CTL0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= inclusion= exclusion= inclusion= exclusion= control (no prime)
modeldata
p(re
spon
se =
prim
e)
} longexpos.
} shortexpos.
orthography
lexical rep.
name
workingmemory
namingpathway
perceptualpathway
Semantic Priming and Response Windows(Draine & Greenwald, 1996; Greenwald, Draine, & Abrams, 1996)
Masked prime word followed by target word
Semantic classification task: Is target pleasant/unpleasant?
Conditions• congruent (FUN-PLAY) vs. incongruent (SICK-PLAY)
• short vs. long prime exposure (“subliminal” vs. “supraliminal”)
• RT to target constrained to fall within temporal response window vs. unconstrained
SimulationDecision pathway makes two-alternativeforced-choice classification
At brief exposure, prime is subliminal
Response window simulated by forcing decisionpathway to closest attractor at time response is required.
decision
perceptual
orthography
semantics
classification
pathway
pathway
PredictionMagnitude of priming effect is larger near center of speed-accuracy curve than near the extremes.
When a forced response is made at a time near the center of the curve, accuracy priming effects will be largest.
100 150 200 250 300 350 4000
0.2
0.4
0.6
0.8
1
cycle
% c
orre
ct
supralim primesublim primenot primed
100 150 200 250 300 350 4000
0.1
0.2
0.3
cycle of forced response
prim
ing
effe
ct
Results
L−C L−I S−C S−I0
0.2
0.4
0.6
0.8
1
L−C L−I S−C S−I0
0.2
0.4
0.6
0.8
1
L−C L−I S−C S−I0
0.2
0.4
0.6
0.8
1
L−C L−I S−C S−I0
0.2
0.4
0.6
0.8
1
no window response windowdata model data model
= long exposure, congruent= long exposure, incongruent
= short exposure, congruent= short exposure, incongruent
accu
racy
L−C L−I B−C B−I570
580
590
600
610
620
L−C L−I B−C B−I200
210
220
230
240
250
260
reac
tion
time
data model
Automatization and Multistep Tasks
Trained network to add two 2-digit numbers in three steps
Each digit represented by pool of units.
4 86 2
1 1 0
1
+
14 86 2
? 1 0
1
+
14 86 2
? ? 0
1
+
?4 86 2
? ? ?
?
+
?step 1 step 2 step 3
hidden units
attractor network
feedforward mapping
copy
result3
result2
result1
carry1
op2-1
op1-2
op1-1
op2-2
carry2
result3
result2
result1
carry1
carry2
result3
result2
result1
carry1
carry2
Sequence of stable states reach awareness; correspond to results of computation
Benefit of attractor net clean up (“projection”)
training set size(% of all problems)
10 20 40 5030
20
40
60
80
100 ProjectionNo projection
Overall performance comparison
Prediction from the Theory: Automatization
Conscious awareness can drop out simply by speeding up processing; doesn’t require a qualitative reorganization of knowledge
Might expect to find evidence for the application of intermediate steps, even without awareness.
E.g., Crutcher (1992) doronico – door – leopard
Explaining Unitary Nature of Consciousness
Often claimed that people can only be aware of one thing at a time.
According to framework, strictly speaking this is false.
Within the output domain of a pathway, a single, coherent representation is selected.
Across pathways, no explicit competition; e.g., stable patterns can be achieved in somatosensory and visual modules simultaneously. But strongly linked domains require mutual consistency.
The subjective impression of unity also arises due to the equating of awareness with verbalization. If we assume there is a module responsible for the verbal expression of ideas, then it is limited to one expression at a time.