Modeling and Neuroscience (or ACT-R and fMRI)
description
Transcript of Modeling and Neuroscience (or ACT-R and fMRI)
Modeling and Neuroscience (or ACT-R and fMRI)
Jon M. FinchamJon M. Fincham
Carnegie Mellon University, Pittsburgh, [email protected]
Jon M. Fincham ACT-R PGSS 2001
Overview
MotivationMotivation Task SpecificsTask Specifics Modeling SpecificsModeling Specifics Experiment ResultsExperiment Results ImplicationsImplications
Jon M. Fincham ACT-R PGSS 2001
“Neuroscience” issues Where does x take place?Where does x take place? What does circuit x do? What does circuit x do? How is x computed?How is x computed?
“Modeling” issues How is x computed? How is x computed? Where does x take place?Where does x take place? What circuit participates in x?What circuit participates in x?
Jon M. Fincham ACT-R PGSS 2001
Modeling & fMRI Issues Computational cognitive modeling provides rich Computational cognitive modeling provides rich
predictions of behavior over time. Can we use the predictions of behavior over time. Can we use the richness of a cognitive model to drive fMRI data richness of a cognitive model to drive fMRI data analysis and if so how do we do it?analysis and if so how do we do it?
How can we use fMRI results to guide How can we use fMRI results to guide development of specific cognitive models and development of specific cognitive models and ACT-R theory in generalACT-R theory in general
Jon M. Fincham ACT-R PGSS 2001
The Task: Tower of Hanoi (of course)
The 5-disk Tower of Hanoi (TOH) task is behaviorally rich planning task
The subgoaling strategy involves varying numbers of planning steps at each move while progressing toward the goal state
ACT-R cognitive model nicely captures behavioral data
Jon M. Fincham ACT-R PGSS 2001
Task Summary: Pre-scan practice
21 pseudo-random problems, classic interface, explicit subgoal posting, mousing
21 pseudo-random problems, grid interface, explicit subgoal posting, mousing
7 problems, grid interface, secondary task, no subgoal posting, 3 button response
Memorize single goal state, 10 simple practice problems
Jon M. Fincham ACT-R PGSS 2001
TOH Classic Interface
Jon M. Fincham ACT-R PGSS 2001
TOH Grid Interface
Jon M. Fincham ACT-R PGSS 2001
The Subgoaling Strategy
1. Select largest out of place disk in current context and destination peg.
2. If direct move, do it and goto step 1. Otherwise, set subgoal to make move
3. If next largest disk blocks destination, select it and other peg & go to step 2.
4. If next largest disk blocks source, select it and other peg & go to step 2.
Jon M. Fincham ACT-R PGSS 2001
TOH 3-tower example move sequence
Plan 3 move sequence (3-C, 2-B, 1-C)
Plan 1 move sequence (2-B)
Plan 1 move sequence (1-B)
Plan 2 move sequence (2-C, 1-A)
Plan 1 move sequence (2-C)
Plan 1 move sequence (1-C)
Plan 1 move sequence (3-C) Goal State
Jon M. Fincham ACT-R PGSS 2001
The Task: TOH in the magnet
One full volume (25 slices) every 4 secondsOne full volume (25 slices) every 4 seconds 16 seconds per move = 4 scans per move16 seconds per move = 4 scans per move 12 20-23 move problems, about 6 minutes 12 20-23 move problems, about 6 minutes
eacheach
Jon M. Fincham ACT-R PGSS 2001
Behavioral Results
1_1c_s 2_2b 3_1b 4_3c 5_1a_s 6_2c 7_1c 8_big1500
2000
2500
3000
3500
4000
4500
3-disk Tower Move Sequence
move_type
Jon M. Fincham ACT-R PGSS 2001
Behavioral Results
Individual Performace on 3-disk Tower
0
1000
2000
3000
4000
5000
6000
m1 m2 m3 m4 m5 m6 m7 m8
Move
s1
s2
s3
s4
s5
s6
s7
s8
Jon M. Fincham ACT-R PGSS 2001
What do we want to see?
How does the brain handle goal processing?How does the brain handle goal processing? Which brain areas are differentially Which brain areas are differentially
responsive to goal setting operations?responsive to goal setting operations? Are there identifiable circuits that Are there identifiable circuits that
collectively implement manipulation of collectively implement manipulation of goals?goals?
Jon M. Fincham ACT-R PGSS 2001
Terminology
BOLD - Blood Oxygenation Level BOLD - Blood Oxygenation Level Dependent response (aka hemodynamic Dependent response (aka hemodynamic response)response)
MR - magnetic resonance, signal measured MR - magnetic resonance, signal measured in the magnetin the magnet
Voxel - approximately cube “point” within Voxel - approximately cube “point” within the brainthe brain
Jon M. Fincham ACT-R PGSS 2001
Where do we begin?
Run model over problem set, collecting Run model over problem set, collecting goal setting event timestampsgoal setting event timestamps
Use goal setting timestamps to generate an Use goal setting timestamps to generate an ideal BOLD-like timeseriesideal BOLD-like timeseries
Jon M. Fincham ACT-R PGSS 2001
ACTR(t) Events and Time Series
Jon M. Fincham ACT-R PGSS 2001
BOLD Response CharacteristicsAdditivity of BOLD Response
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
0 5 10 15 20 25
Time
Cumulative
Response 1
Response 2
Response 3
Jon M. Fincham ACT-R PGSS 2001
Identifying a responsive voxel
Model MR signal as a function of the ACT-R Model MR signal as a function of the ACT-R generated time seriesgenerated time series
MR(t) = BMR(t) = B00 + B + B11*trial(t) + B*trial(t) + B22*ACTR(t) + *ACTR(t) + (t)(t) Ignore error trials and immediate successorsIgnore error trials and immediate successors
Run regression for every one of the 25x64x64 Run regression for every one of the 25x64x64 voxelsvoxels
Result is a beta map for each regressorResult is a beta map for each regressor
Jon M. Fincham ACT-R PGSS 2001
Group Analysis
Morph each brain into a reference brainMorph each brain into a reference brain
Voxel-wise 2-tailed t-test of HVoxel-wise 2-tailed t-test of H00: B: B22 = 0 across = 0 across
subjectssubjects
Jon M. Fincham ACT-R PGSS 2001
Analysis Summary
Within subject voxel-wise regression of MR Within subject voxel-wise regression of MR signal against ACT-R generated time seriessignal against ACT-R generated time series MR(t) = BMR(t) = B00 + B + B11*trial(t) + B*trial(t) + B22*ACTR(t) + *ACTR(t) + (t)(t) Ignore error trials and immediate successorsIgnore error trials and immediate successors
Voxel-wise 2-tailed t-test of HVoxel-wise 2-tailed t-test of H00: B: B22 = 0 across = 0 across
subjectssubjects Threshold at p<0.0005 and contiguity of 8 voxelsThreshold at p<0.0005 and contiguity of 8 voxels
Jon M. Fincham ACT-R PGSS 2001
TOH Activation Map (p < 0.0005, contiguity = 8)
R L
Jon M. Fincham ACT-R PGSS 2001
Premotor & Parietal activity increase parametrically with number of planning steps
Jon M. Fincham ACT-R PGSS 2001
Premotor & Parietal activity increase parametrically with number of planning steps
Jon M. Fincham ACT-R PGSS 2001
Premotor & Parietal activity increase parametrically with number of planning steps
Jon M. Fincham ACT-R PGSS 2001
Prefrontal - Basal Ganglia - Thalamic Circuit
Jon M. Fincham ACT-R PGSS 2001
Prefrontal - Basal Ganglia - Thalamic Circuit
Jon M. Fincham ACT-R PGSS 2001
Prefrontal - Basal Ganglia - Thalamic Circuit
Jon M. Fincham ACT-R PGSS 2001
Prefrontal - Basal Ganglia - Thalamic Circuit
Jon M. Fincham ACT-R PGSS 2001
PFC -Basal Ganglia -Thalamus
Striatum = Pattern Striatum = Pattern Matching & conflict Matching & conflict resolution?resolution?
Result gates thalamus Result gates thalamus to update buffers?to update buffers?
Cortex
Thalamus
GP Striatum
Jon M. Fincham ACT-R PGSS 2001
Summary of findings so far... Move planning activity in parietal and premotor Move planning activity in parietal and premotor
areas varies parametrically with number of areas varies parametrically with number of planning stepsplanning steps
PFC-Basal Ganglia-Thalamic circuit does not vary PFC-Basal Ganglia-Thalamic circuit does not vary parametrically with number of planning steps but parametrically with number of planning steps but shows significant BOLD response during high shows significant BOLD response during high planning moves onlyplanning moves only
Suggests PFC becomes engaged when sequencing Suggests PFC becomes engaged when sequencing of multiple moves is requiredof multiple moves is required
Jon M. Fincham ACT-R PGSS 2001
What can we conclude about the model? Subjects are bypassing subgoaling Subjects are bypassing subgoaling
procedure for 2-tower subproblemsprocedure for 2-tower subproblems Setting a goal “move disk 1 to opposite of Setting a goal “move disk 1 to opposite of
where disk 2 goes” where disk 2 goes”
Now we can use GLM model comparison Now we can use GLM model comparison techniques to confirm best fitting models...techniques to confirm best fitting models...
Jon M. Fincham ACT-R PGSS 2001
What can we conclude about ACT-R? Nothing…….yet.Nothing…….yet. Goal manipulation does seem to predict Goal manipulation does seem to predict
brain activity in the “right” places, butbrain activity in the “right” places, but Need to run other studies in different Need to run other studies in different
domains (and different models) to gain domains (and different models) to gain confidence in our label of “goal processing” confidence in our label of “goal processing” circuitrycircuitry
Jon M. Fincham ACT-R PGSS 2001
What have we learned so far?
Applying cognitive modeling to the Applying cognitive modeling to the neuroimaging domain is feasible: models can neuroimaging domain is feasible: models can inform analysisinform analysis
fMRI data can inform models fMRI data can inform models fMRI data can inform architecturefMRI data can inform architecture Symbiotic relationship exists between Symbiotic relationship exists between
modeling and fMRImodeling and fMRI What else?What else?
Jon M. Fincham ACT-R PGSS 2001
What else can we examine?
+goal>, +retrieval>, +visual>, +aural>, +goal>, +retrieval>, +visual>, +aural>, +manual>, +manual>,
Number of elements in goalNumber of elements in goal Number of full buffersNumber of full buffers
Jon M. Fincham ACT-R PGSS 2001
Thank you!