Computational Analysis of Motor Learning

advertisement
Computational Analysis of
Motor Learning
Motor Learning
Three paradigms
Force field adaptation
Visuomotor transformations
Sequence learning
Does one term (motor learning) fit all?
How to determine similarities/differences?
Acquisition: On-line vs. KR feedback
Generalization: Transfer
Motor Learning
Three paradigms
Force field adaptation
Visuomotor transformations
Sequence learning
Does one term (motor learning) fit all?
Neural systems: Do these tasks engage common regions?
Motor Learning
Three paradigms
Force field adaptation
Visuomotor transformations
Sequence learning
Does one term (motor learning) fit all?
Neural systems: Do these tasks engage common regions?
Test case: Cerebellum
Force field adaptation is impaired in patients with cerebellar
degeneration and not basal ganglia degeneration.
Smith and Shadmehr, 2005
Prism adaptation impairment in patient with bilateral
cerebellar stroke.
Control Participant
Cerebellar Stroke
Martin et al., 1996
Sequence learning is absent (ipsilesional) in patients with
unilateral cerebellar stroke.
S S S R S S S R
Gomez-Beldarrain et al., 1998
Classical conditioning of eyeblink response
Cerebellar lesions selectively abolish the learned
response.
McCormick et al., 1984; many others
Model system of motor learning
1. Dissociation of performance
and learning.
2. Many species.
3. Specification of US and CS
pathways.
-- US, CS simulations.
-- Reversible lesions
4. Genetic manipulations.
Cerebellar lesions selectively abolish the learned
response.
But not all forms of classical conditioning.
Pre-training lesions
McCormick et al., 1984; many others
Specifying domain of function.
Task domain:
Cerebellum for motor learning.
Computation:
Cerebellum for learning precise
timing between stimulus events.
Specifying domain of function.
Task domain:
Cerebellum for motor learning.
Computation:
Cerebellum for learning precise
timing between stimulus events.
Timing hypothesis accounts
for dissociation of heart rate
and eyeblink responses.
Lesions of the cerebellar hemisphere do not abolish
eyeblink response but do disrupt the adaptive timing.
Prelesion
Postlesion
Trained with
200 ms ISI
Trained with
500 ms ISI
Perrett et al., 1993
Multiple levels of learning.
Simple associations at DCN
Precise timing from cerebellar
cortex to flexibly make
response adaptive.
Motor Learning
Three paradigms
Force field adaptation
Visuomotor transformations
Sequence learning
Neural systems: Do these tasks engage common regions?
Revisit role of cerebellum:
Do these tasks require precise timing?
Sequence learning as test case.
Stimuli appear at one of
four positions.
Press response key in
corresponding position.
Stimuli follow sequence or
are chosen at random.
Sequence learning as test case.
Stimuli appear at one of
four positions.
Press response key in
corresponding position.
Stimuli follow sequence or
are chosen at random.
Learning series of spatial associations.
Stimuli, responses lack precise timing.
Transfer indicates effector-independent learning.
Computational analysis:
Cerebellum not essential for sequence learning.
Avalanche of patient and imaging studies with SRT task
central question: Is Structure X involved in learning?
Extensive focus on basal ganglia and cerebellum given hypothesized
role in skill, procedural learning, and automaticity.
Patient Groups
Basal ganglia
Parkinson’s:
Focal BG lesions
Cerebellum
Degree of Learning Compared to Controls
Normal Attenuated None
2
2
0
4
1
1
1
0
5
Patient Groups
Basal ganglia
Parkinson’s:
Focal BG lesions
Cerebellum
Imaging results
Degree of Learning Compared to Controls
Normal Attenuated None
2
2
0
4
1
1
1
0
5
Directional Change in Activation with Learning
Increases No Change Decreases
Basal Ganglia
5
6
0
Cerebellum
0
6
4
Patient Groups
Basal ganglia
Parkinson’s:
Focal BG lesions
Cerebellum
Degree of Learning Compared to Controls
Normal Attenuated None
2
2
0
4
1
1
1
0
5
Imaging results
Directional Change in Activation with Learning
Increases No Change Decreases
Basal Ganglia
5
6
0
Cerebellum
0
6
4
If both BG and Cerebellar lesions impair SRT learning, why are
imaging results so different?
If cerebellar lesions are so devastating to learning, why decrease with
learning, esp. when motor cortical areas show increase w/ learning?
Sequence learning as test case.
Performance-based hypothesis:
Cortical-cerebellar
interactions to maintain
S-R mapping.
Non-cerebellar systems for
forming sequential associations.
Patient deficit: Poor sequence learning because noisy S-R codes
provide weak input for associative mechanisms.
Prediction: Deficit in SRT learning will be reduced in patients
with cerebellar lesions if S-R coding requirements
are minimized.
Prediction: Deficit in SRT learning will be reduced in patients
with cerebellar lesions if S-R coding requirements
are minimized.
Reach to target
location
Reach to target associated
with central color
Successive targets
follow 8-element
sequence or selected
randomly.
Symbolic cues tax
S-R coding system
(e.g., Wise premotor)
Patients selectively impaired with symbolic cues.
Consistent with performance problem rather
than sequence learning per se.
Experiment 2:
Use more traditional keypressing SRT task
Compare with patient control: PD patients
Experiment 2:
Use more traditional keypressing SRT task
Compare with patient control: PD patients
Ataxia group
shows selective
impairment with
symbolic cues.
Controls and PD
learn with both
cues.
Results bring together patient and imaging work.
Lesion
All human cerebellar studies have used symbolic cues.
One monkey lesion study used direct cues: normal performance
Results bring together patient and imaging work.
Lesion
All human cerebellar studies have used symbolic cues.
One monkey lesion study used direct cues: normal performance
Imaging
Consistent with imaging studies showing reduction in cerebellar
activation over the course of learning.
Demands on maintaining S-R mapping will decrease over time
as mapping becomes well-learned.
Directional Change in Activation with Learning
Increases No Change Decreases
SRT learning
0
6
4
Results bring together patient and imaging work.
Lesion
All human cerebellar studies have used symbolic cues.
One monkey lesion study used direct cues: normal performance
Imaging
Consistent with imaging studies showing reduction in cerebellar
activation over the course of learning.
Demands on maintaining S-R mapping will decrease over time
as mapping becomes well-learned.
Directional Change in Activation with Learning
Increases No Change Decreases
SRT learning
0
6
4
Prediction: Weak cerebellar/prefrontal activation with direct cues.
Motor Learning
Cerebellar role in learning
Sequence learning
Indirect (no timing)
Force field adaptation
likely (on-line timed error signal)
Visuomotor transformations
???
Motor Learning
Cerebellar role in learning
Sequence learning
Indirect (no timing)
Force field adaptation
likely (on-line timed error signal)
Visuomotor transformations
???
Imaging and patient work suggests yes.
Working memory account:
Two S-R maps required in transformed environment
S-R Map 1: Old Map
S-R Map 2: Hypothesis of New Map
Motor Learning
Cerebellar role in learning
Sequence learning
Indirect (no timing)
Force field adaptation
likely (on-line timed error signal)
Visuomotor transformations
???
Compare single- and multi-step transformations.
Single: 25 deg displacement in one step
Multi: 5 deg every 20 trials
Predictions?
Classic model of cerebellum and error detection and
correction.
Parallel fibers: Simple spikes indicate context.
Climbing fibers: Complex spikes indicate error.
Complex spike activity leads to weight
change.
“Constant”
somatosensory input
that is either expected
or unexpected
Unexp. Exp.
Unexp
Unexp. Exp.
Unexp
Exp
Unexp
Errors of Omission and Commission
Yes
Expected Yes
No
Actual
No
Correct
Commission
Errors of commission are well-timed.
Errors of omission lack precise timing.
Omission
Correct
Download