Download Bischoff_Thesis - The USC Brain Project

Bischoff(Grethe)-Arbib Basal
Ganglia Modeling
Presented by James Bonaiuto
2/11/2007
Amanda Bischoff (Grethe)’s
Thesis
• Models the basal ganglia (and some
cortical areas) in three tasks:
– Elbow flexion-extension
– Reciprocal aiming
– Sequential arm movements
• Dopamine levels were modified to model
the effects of Parkinson’s
2/11/2007
Hypothesis on Basal Ganglia
Function
• Basal Ganglia
– Indirect pathway –
movement inhibition
– Direct pathway –
provides next sensory
state to cortex
• Cortex
– Preparatory areas –
project to indirect path
– Movement-related
areas – project to
direct pathway
2/11/2007
Model Overview - Cortex
• Pre-SMA
– Projects sequential information to SMA and indirect pathway
• SMA
– Contains information on the overall sequence
– Keeps track of which movement is next
– Project current movement to MC and direct pathway of basal
ganglia
– Project next movement to premovement population in MC and
indirect pathway of basal ganglia
• Motor Cortex
– Carries out motor command
– Handles fine-tuning of movement
– Projects motor parameters to brainstem and direct pathway of
basal ganglia
2/11/2007
Next Sensory State Information
• Why aren’t the basal ganglia responsible for
movement initiation?
– Crutcher & Alexander (1990) – movement related
putamen neurons fire an average of 33 ms after the
onset of a movement (after activation of MC – 56ms
later, and SMA – 80 ms later)
– Mink & Thach (1991b) – movement-related activity in
GPe and GPi is also late
– Turner & Anderson (1997) – GP neurons rarely
change discharge before activity of agonist muscles
2/11/2007
Basic Model
• Segregated direct
(movement)/indirect
(preparation) pathways
• Neat modeling trick:
– To model up/down states of
putamen neurons, the time
constant is a sigmoid of the
membrane potential
– Same trick is used later to
slowdown the cortex in the
absence of dopamine
2/11/2007
Elbow Flexion-Extension
2/11/2007
Elbow Flexion-Extension - Results
<Demonstration>
2/11/2007
Reciprocal Aiming
• Winstein et al. (1997) – Stylus tapping between two
targets of varying sizes
• Fitt’s Law – speed/accuracy tradeoff
– ID=log2(2A/W)
– MT=a+bID
• Parkinson’s patients
– Slower overall time
– Constrained trajectory
– Reached to smaller area of target
• Predictions:
– Slower speed is due to inability of BG to release inhibition of
movement
– Decrease in SMA and MC activity causes reduction in speed and
variation of movement
2/11/2007
Reciprocal Aiming - Model
• Input: target positions in joint
space
– Problem when targets
overlap in joint space
• SMA_INH prepares
upcoming movement – BG
inhibits before appropriate
– WTA– only fires in relation to
movement in preparation
• SMA_MVT receives info from
both targets
– Inhibition from SMA_INH –
only responds to current
target
• MC_MVT
2/11/2007
– Encodes joint coordinates converted to Cartesian
space
– Movement time calculated
from firing rate
Reciprocal Aiming - Results
• Normal - Qualitatively similar to Winstein
et al.’s (1997) control data
• 50% Dopamine
– No contact with target, no pause between
movements
• Because neural part of model taking less time
than arm
– Hypothesis: slowdown in putamen function
may cause slowdown in cortex too
• Changed time constants of SMA and MC to
depend on dopamine level
• With dopamine depletion – takes longer for
neurons to reach maximum and maximum is
less than with dopamine (because of longer time
constant)
• Reduction in MC firing rates causes delays
between movements
• Caused restricted arm trajectory – lower velocity
2/11/2007
Reciprocal Aiming Results
SMA-Proper
2/11/2007
Motor Cortex
Reciprocal Aiming Results
Putamen
2/11/2007
GPe
STN
GPi
SNc
Reciprocal Aiming Results
Normal
2/11/2007
50% Dopamine
20% Dopamine
Sequential Arm Movements
• Extends SMA module for a sequence of
three movements
• Tanji & Shima (1994) – SMA neurons
selective for sequence order, others
selective for movement no matter where it
was in a sequence
• Tanji & Mushiake (1996) - Pre-SMA active
for visual stimuli – indicate sequence to be
performed
2/11/2007
Sequential Arm Movements Model
• Pre-SMA
– Now selective for different sequence
permutations
• SMA
– New population selective for
different sequence permutations and
subsequences
• After the current movement begins,
SMA_INH primes SMA_MVT for the
next movement
• MC_MVT needs to reach a
threshold firing rate to produce
target for movement generator
• Hardcoded relationships between
SMA_SEQ, SMA_MVT and
SMA_INH
2/11/2007
Sequential Arm Movements Results
SMA-Proper
Motor Cortex
•
Seq123 and seq12 active until target 1 reached
•
Seq12 primes target 2 neurons in SMA_PROPER_INH and seq23
•
Target 1 reached – seq23 reaches full activation
•
Seq23 primes target 3 neurons in SMA_PROPER_INH
•
Drop dopamine - seq123 is active longer
2/11/2007
– MC_MVT peaks for each movement lower than for previous one - each movement depends on activation
from previous movement
Sequential Arm Movements Results
Putamen
2/11/2007
GPe
STN
GPi
SNc
Sequential Arm Movements Results
• Reduce dopamine
Normal
50%
Dopamine
20%
Dopamine
2/11/2007
– Beginnings of pause between each
submovement
– Akinesia - took longer to initiate 1st
movement
– Bradykinesia – each movement take
longer and longer
• Indirect pathway is overactive
(inhibits motor programs), direct
pathway is less capable of
responding to current motor
command
• Slower time constant and higher GPi
inhibition -> SMA doesn’t know
status of current motor program so
doesn’t command the next
movement
Discussion
• Can the same model do all three tasks?
– Reciprocal aiming and flexion-extension can be cast
as 2 movement sequences
– Requires new weights for the SMA_SEQ12 and
SMA_SEQ21 populations
– How can these weights be learned?
• The future work section lists the inclusion of
cortico-STN projections
– The GPR model includes these, but has an opposite
take on the basal ganglia function (action selection)
– Are these views reconcilable?
2/11/2007