Gluck, Myers, & Shohamy. Rutgers-Newark Neuroscience. 1. Why are BG
important for category learning? Perhaps feedback is key. 2. Further
manipulation of ...
3.2 Gluck
When are Parkinson’s Patient’s Impaired (or Not) on Category Learning
When are Parkinson’s Patients Impaired (or Not) at Category Learning
Gluck, Myers, & Shohamy. Rutgers-Newark Neuroscience
1. Why are BG important for category learning? Perhaps feedback is key. 2. Further manipulation of task variables
M. Gluck, C. Myers, D. Shohamy Rutgers-Newark Center for Molecular & Behavioral Neuroscience
Goal of Study
Converging Evidence Suggests BG Important for Feedback Learning 1. Electrophysiology: BG modify responses based on (rewarding?) feedback (e.g. Ljunberg et al.,
Compare Parkinson’s patients on probabilistic category learning under feedback and no-feedback (“observational”) training
1992;Schultz, 1997)
Prediction
2. fMRI: BG active during feedback learning, not observational learning (Poldrack et al., 2001). 3. Neuropsych: Parkinson’s patients impaired on some feedback learning tasks (Knowlton et al., 1996; Myers et
Parkinson’s patients will be • impaired when learning is feedback-based • not impaired when learning is observational.
al., 2003).
•
However, necessity of BG for feedback learning not demonstrated directly.
“Mr. Potatohead”
Feedback Condition
Hat
.8
.2
Glasses
.6
.4
Moustache
.4
.6
Bowtie
.2
.8
A.
B.
Which flavor do you think he wants?
Vanilla
Correct!
Features on Mr. Potatohead predict category outcome (vanilla or chocolate) probabilistically
1
3.2 Gluck
Observational Condition
Subjects: PD & Control Age PD FB CON
PD OB CON
Press “next” to see another customer
Education
61.3
16.6
(8.4)
(2.4)
MMSE
H-Y
Years PD
29.2
2.3
6.2
(0.8)
(0.8)
(3.2)
59.0
17.0
29.8
N/A
N/A
(6.4)
(2.4)
(0.4)
64.5
15.9
(6.0)
29.0
2.1
5.4
(1.2)
(0.7)
(4.7)
64.1
16.9
29.0
N/A
N/A
(6.0)
(3.3)
(2.3)
(0.9)
FB = feedback task, OB = observational task, MMSE = Mini Mental State Exam; age, education and PD duration in years. SD in parentheses.
• Non-demented and non-depressed. • Intact cognitive function. • Tested on medication.
Results for Mr. Potatohead
Results for Mr. Potatohead
% correct
% correct
100
100
Controls PD
90
Controls PD
90
80
80
70
70
60
60
50
50
40
40
Feedback
Feedback
Consistent with prior studies: PD impaired with feedback learning
Learning strategies Math analyses determine fit of individual data to models of learning strategies Prior studies: Most subjects use specific subset of strategies under feedback conditions (Gluck et al., 2002)
Do PD and controls use same strategies with observational learning?
Observational
PD not impaired learning same task with ‘observational’ learning
Learning Strategies Prior studies: 3 Main Learning Strategies 1.
Multi-cue (learn all 4 cues;optimal)
2.
Singleton (learn 4 single-cue patterns)
3.
One-cue (respond based
on one cue)
For Details See:
Gluck, M. A., Shohamy, D., & Myers, C. E. (2002). How do people solve the “weather prediction” task?: Individual variability in strategies for probabilistic category learning. Learning and Memory. 9. 408-418.
2
3.2 Gluck
Feedback Learning Strategies 100
singleton
80 70
% subjects
Almost NONE of the subjects in any group use the strategies defined earlier.
multi cue
90
Observational Learning: No Strategies
one cue
60 50
Many subjects appear to learn a RULE, responding correctly to a subset of patterns/exemplars.
40 30 20 10 0
Controls
PD
• Almost all subjects use one of previously defined strategies. • PD show different pattern of strategy use than controls
Subjects learn the task in a qualitatively different manner under feedback vs. observational conditions.
Feedback vs. Observational Differentially Recruit Striatum and MTL
Summary PD impaired on PCL only when learning is feedback-based, not observational.
Direct comparison:
Feedback vs. observational learning invoke different strategies in both controls and PD.
FB > Obs: Striatum and thalamus, midbrain Obs > FB: MTL and PFC
Consistent with role for BG in modifying behavior based on response-contingent feedback as suggested by our prior imaging study with R. Poldrack (next slide..)
Poldrack et al., 2001, Nature
?:
Does this suggest a double dissociation
Feedback BG damage (PD) MTL Damage (Amn)
?:
Does this suggest a double dissociation
Feedback
Observ.
Impaired OK ?OK? ?Impaired?
BG damage (PD) MTL Damage (Amn)
Observ.
Impaired OK ?OK?* ?Impaired?
* Prior data suggested early MTL learning OK (Knowlton, Squire, & Gluck, 1994; Knowlton, Mangels, & Squire, 1996)
3
3.2 Gluck
Exp 1. Weather Prediction
Exp. 2. Mr. Potatohead
Hopkins, R., Myers, C., Shohamy, D., Grossman, S., & Gluck, M. (In press). Impaired category learning in hypoxic subjects with hippocampal damage. Neuropsychologia, % Correct 100
80
Control l
l l
90
Control
l l
80
l
60
m
70
l
60
m
m m
m
2
3
Hypoxic
50
m m
m
40
50
Hypoxic
4
60
l
l
m
m
2
3
Hypoxic
50 20
30
Trials
In contrast to Knowlton et al. (1994, 1996) we find MTL/Amnesic deficit both early and late in training. Why?: These hypoxic amnesics have verified MTL damage, while earlier studies had mixed etiologies with broader damage.
Control
l
80 70
m
m
m 10
Blocks (of 50 trials)
90 l l
l
l
70
50 1
100
Control
90 80
l
l 70
% Correct
100
100
90
Hopkins et al. Neuropsychologia, In press. % Correct
% Correct
l m
60
l m
m
m
Hypoxic
m
50 1
4
10
20
Blocks (of 50 trials)
30
40
50
Trials
Same result with Mr. Potatohead task.
No clean double dissociation: Comparative Strategy Analyses Feedback
Hopkins et al. Neuropsychologia, In press.
(A) Weather Task (B) Mr. Potatohead % Subjects 100 80
% Subjects C
100
C
H
80
H
60
60
40
40
20
20
0 Multi-cueOne-cue Sing.
Best-Fit Strategy
0 Multi-cueOne-cue Sing.
Best-Fit Strategy
Event-related fMRI of Weather Prediction
BG damage (PD) MTL Damage (Amn)
Impaired Impaired
Observ. OK ??
But..consistent with earlier imaging and modeling:
BG and MTL both important for PCL but in different ways, and at different times during learning (early MTL, late BG)
The Computational Model:
Poldrack, Clark, Pare-Blageov, Shohamy, Creso-Moyano, Myers, & Gluck, Nature (2001)
Response basal ganglia
Feedback hippocampus
Recode stimulus representations
Medial Temporal Lobe
cortex
(Gluck & Myers, 1993)
Weather cards
Weather cards
Output signal
Training signal
BG/Caudate nucleus
Results: Early MTL activity, late BG activity
4
3.2 Gluck
The Computational Model:
The Computational Model:
Early in Training Response
Late in Training
Feedback
Response hippocampus
basal ganglia
Recode stimulus representations cortex
(Gluck & Myers, 1993)
Weather cards
Weather cards
Feedback
basal ganglia
Output signal
hippocampus
Output signal
Recode stimulus representations New stimulus representations
Training signal
Weather cards
The Computational Model:
Training signal
(Gluck & Myers, 1993)
Weather cards
When are Parkinson’s Patient’s Impaired (or Not) on Category Learning Gluck, Myers, & Shohamy. Rutgers-Newark Neuroscience
Late in Training Response basal ganglia
1. Why are BG important for category learning? Perhaps feedback is key.
Feedback hippocampus
Recode stimulus representations cortex
(Gluck & Myers, 1993)
Weather cards
Weather cards
Output signal
Training signal
2. Further manipulation of task variables to better understand functional roles of BG and MTL in category learning
“Slots”
Myers et al. (2003) CNS Poster 1. Are PD simply slower at probabilistic category learning (PCL), or is there a qualitative difference in how they approach PCL, relative to controls? 2. Is the PD deficit unique to the (very difficult) weather task, or does it extend to other (easier) probabilistic category tasks?
Example “white coins” trial: P(white)=.8
P(white)=.2
P(black)=.2
P(black)=.8
Cue 1
or
Cue 2
or
Cue 3
or
Example “black coins” trial:
Part 2 Instructions are the same as for Part 1, but you will be using a NEW slot machine. Combinations that give white coins and black coins are now different. When ready to start Part 2, press the "White Coin" key
5
3.2 Gluck
Slots: Simpler PCL
Slots: PD vs. Matched Controls
Stimuli
Acquisition % Correct
Three independent cues.
Each pattern deterministically associated with each outcome * subject potentially achieve 100% correct (vs. 67% WP) Fewer solution strategies than in “weather” task: * no “singleton” patterns
Slots Strategy Analysis:
40 20
MultiCue
0 1 2 3 Best-Fit Strategy
Mix
40
20
20
1
2
3 4 5 6 7 8 Blocks (of 10 trials)
9 10
1
2
3 4 5 6 7 8 Blocks (of 10 trials)
9 10
Thus, PD deficit observed in “weather” task may depend on specific features of that task, including overall difficulty, need to encode configural cues, pattern-response conflict, etc., rather than reflecting a general PD deficit in probabilistic classification learning. Future: manipulation of these variables independently
% Subjects 60
PD
0
40
• PD are not impaired at acquisition.
Reversal Control
60
Slots: General Discussion
PD shift, Controls Reverse
20
PD 80
60
• Reversal: PD are not different from controls.
Acquisition
40
Control
PD
80
• Acquisition: PD are not different from controls.
Transfer: Unsignaled reversal
% Subjects 60
Reversal % Correct 100
Control
100
Each cue probabilistically associated with each outcome
MultiCue
1 2 3 Best-Fit Strategy
Mix
• Acquisition: Most subjects use a one-cue strategy • Reversal: All but 2 controls best-fit by same strategy as in
acquisition; all but 2 PD shift from a one-cue strategy to another (different) one-cue strategy. => Whereas Controls keep (but reverse) their earlier strategy, PD “avoid” reversal by shifting to a new, equally-predictive cue.
• No PD impairment on reversal. PD “avoid” reversal by shifting to another, equally-effective strategy Lack of PD impairment on this shifting is consistent with the general lack of impairment by medicated PD patients on other intradimensional shift tasks (e.g. Downes et al., 1989; Gauntlett-Gilbert et al., 1999).
Bibliography Gluck, M., Shohamy, D., & Myers, C. (2002). How do people solve the "weather prediction" task? Individual variability in strategies for probabilistic category learning. Learning and Memory, 9(6), 408-418. Hopkins, R., Myers, C., Shohamy, D., Grossman, S., & Gluck, M. (in press). Impaired category learning in hypoxic subjects with hippocampal damage. Neuropsychologia, to appear Myers CE, Shohamy D, Gluck M, Grossman S, Kluger A, Ferris S, Golomb J, Schnirman G, Schwartz R. (2003). Dissociating hippocampal vs. basal ganglia contributions to learning and transfer. Journal of Cognitive Neuroscience, 15:185-193.
The End
Poldrack RA, Clark J, Pare-Blageov J, Shohamy D, Creso Moyano J, Myers C, Gluck MA (2001) Interactive memory systems in the human brain. Nature 414:546-550. Shohamy, D., Myers, C., Onlaor, S., & Gluck., M. (2003). The role of the basal ganglia in category learning: How do patients with Parkinson’s disease learn? Manuscript under editorial review.
6
3.2 Gluck
Slots Methods Subjects • 13 individuals with mild-to-moderate idiopathic PD - 10 males, 3 females; mean age 62.4 years; mean education 16.8 years. - All tested on dopaminergic medication; clean of other medication including anticholinergics; screened for depression and dementia. • 13 healthy controls - 6 males, 7 females; mean age 63.0 years, mean education 15.9 years. Neither age nor education differed significantly from the PD group (age: t(24)=0.31, p>.500; education: t(24)=0.75, p-.462). - Screened for absence of any neurological or psychiatric disorder, including depression; free of any medication that could impair cognition.
Procedure • Task and data analysis same as in Experiment 1.
7