Transcript Background Presentation

Functions of Distributed Plasticity in a
Biologically-Inspired Adaptive Control
Algorithm: From Electrophysiology to
Robotics
University of Edinburgh
University of Sheffield
University of the West of England
1. Background to project
AwayDay 2005
Slide No 2
•
•
•
In some respects animal movements better than robot movement
Could in part be due to characteristics of biological control algorithms
Which region of the brain particularly concerned with skilled movement?
AwayDay 2005
Slide No 3
Cerebellum
• located at base of brain (here a human brain)
• looks like a small version of overlying cerebral cortex?
• cerebellum = ‘little brain’
AwayDay 2005
Slide No 4
Cerebellar Function
•
Clinical and experimental
observations of cerebellar
damage
•
Does not cause paralysis, but
makes many movements
inaccurate, slow and
uncoordinated
•
Similar to effects of alcohol:
tests for intoxication may
resemble clinical test for
cerebellar impairment
AwayDay 2005
Slide No 5
• Conclusion: cerebellum is particularly associated with
those features of movements that distinguish animals
from robots
• Framework of project: to investigate whether there
are features of cerebellar control that are likely to be
of interest to robotics
AwayDay 2005
Slide No 6
1. Framework: is cerebellar control of interest to robotics?
2. Problem
AwayDay 2005
Slide No 7
Cerebellar Cortex
• Adjacent to and connected with the the brainstem
• Has its own cortex (= rind)
AwayDay 2005
Slide No 8
Cerebellar Cortex
• Small number of cell
types in cerebellar
cortex
•
AwayDay 2005
Connected to form a
distinctive microcircuit
Slide No 9
Cerebellar Microcircuitry
• Classic work published in
1967
• Investigated anatomy and
electrophysiology of
microcircuit
• Same basic circuit repeated
many times (hence
“neuronal machine”)
• Important: half the cells in
the entire brain are in the
cerebellum
AwayDay 2005
Slide No 10
Mossy Fibres
Idea of Cerebellar ‘Chip’
• Structure of cerebellar cortex
is very uniform over its entire
surface
• Different regions have
different inputs and outputs,
(microzones) but same basic
organisation
• Gives rise to idea of
cerebellar chip: ~5000, each
with its own particular
connections.
AwayDay 2005
Slide No 11
Choose Your Task
• Consequence of this arrangement: all motor tasks
using the cerebellum employ the same basic
cerebellar algorithm
• The investigator can therefore choose the most
‘appropriate’ motor task
• In our case, control of the vestibulo-ocular reflex
(VOR)
AwayDay 2005
Slide No 12
Vestibulo-Ocular Reflex (VOR)
• Vision is degraded if the
image moves (‘slips’) too
much across the retina
• Retinal slip would be
produced by movements of
the head, such as occur in
locomotion
• The VOR acts to counterrotate the eyes to prevent
retinal slip, i.e. to maintain
stable gaze
• Usually not aware when we
use it
AwayDay 2005
Slide No 13
VOR Control: Basic Circuit
Semicircular
canals
Primary
Vestibular
Neurons
Secondary
Vestibular
Neurons
Ocular
Motor
Neurons
Extraocular
Muscles
• Input from vestibular position,
senses head movement
• Passed to interneurons in vestibular
nuclei (secondary vestibular
neurons)
• Thence to motor neurons that control
the eye muscles
• This circuit in brainstem (just below
cerebellum)
AwayDay 2005
Slide No 14
VOR Control: Cerebellum
Flocculus
head
velocity
Brainstem
Retinal slip
motoneuron
firing
Eye
Muscles
Orbital
Tissue
eye
velocity
• Cerebellar flocculus receives information about
– Head velocity
– Eye movement commands
– Retinal slip
• Projects back to brainstem
AwayDay 2005
Slide No 15
VOR Control: Generalised Version
head
velocity
reference
r(t)
AwayDay 2005
Flocculus
and
Brainstem
Controller
motoneuron
firing
command
u(t)
Eye
Muscles
Orbital
Tissue
Plant
eye
velocity
output
y(t)
Slide No 16
Not Feedback Control
reference
r(t)
Controller
command
u(t)
X
Plant
output
y(t)
Sensor
• Retinal slip signal is delayed by 100 ms (visual processing)
• Feedback control would become unstable at ~ 2.5 Hz, yet VOR
operates up to ~25 Hz
• Feedback control not suitable
AwayDay 2005
Slide No 17
Control Method: Open-Loop
reference
r(t)
Inverse
Plant
Model P-1
command
u(t)
Plant
P
output
y(t)
• If feedback not available, then open-loop control must be
used
• If reference signal is desired output, then the controller
becomes an inverse model of the plant (‘plant
compensation’)
AwayDay 2005
Slide No 18
Adaptive Control
desired
output
r(t)
Inverse
Plant
Model P-1
command
u(t)
training signal
Plant
P
output
y(t)
Sensor
• How can we be sure the inverse plant model is accurate?
• Requires constant calibration – ‘adaptive control’
• Use information about system output for learning, rather than online control
AwayDay 2005
Slide No 19
VOR Equivalent
head
velocity
Brainstem
Flocculus
motoneuron
firing
Eye
Muscles
Orbital
Tissue
eye
velocity
retinal
slip
• Available training signal is retinal slip, known to be sent to the flocculus
• Consistent with flocculus being the adaptive part of the controller
• Consistent with e.g. lesion evidence that VOR adaptation is lost after
floccular inactivation
AwayDay 2005
Slide No 20
Why VOR Calibration?
1. Well-defined adaptive control problem
2. Eye movements are relatively simple
– single joint instead of up to ~6 joints in finger movements
– constant load
3. Great deal known about underlying circuitry
4. Well established cerebellar involvement
AwayDay 2005
Slide No 21
1. Framework: is cerebellar control of interest to robotics?
2. Problem: adaptive calibration of VOR
3. Approach: multidisciplinary
AwayDay 2005
Slide No 22
Multidisciplinary Approach
• Modelling
– (theoretical neuroscience, Sheffield)
• Electrophysiology
– (experimental neuroscience, Edinburgh)
• Robotics
– (University of the West of England, Bristol)
AwayDay 2005
Slide No 23
General Modelling Task
• Devise a working algorithm that connects the microcircuit to the
behavioural competence
• Obeys known anatomical and physiological constraints
AwayDay 2005
Slide No 24
Cerebellar Modelling
• Cerebellar microcircuit has
been extensively modelled,
starting with classic work of
Marr (1969) and Albus (1971)
• Here in more modern form of
the adaptive filter
AwayDay 2005
Slide No 25
Specific Modelling Problem
•
Extensive experimental work shows that in VOR
calibration there are TWO sites of plasticity
1. In cerebellar cortex, as predicted by adaptive filter models
2. In the brainstem
•
AwayDay 2005
What are the computational advantages of this
distributed plasticity?
Slide No 26
Electrophysiology: Problem
Mayank B Dutia
Centre for Integrative Physiology
University of Edinburgh
AwayDay 2005
• What are the learning
rules underlying
brainstem plasticity?
• Existence known for ~20
years, rules yet to be
identified
• Critical for understanding
computation significance
Slide No 27
Electrophysiology: Technique
Rostral
Midline
Medial
Vestibular
Nucleus
• Record from neurons in
slices through brainstem
• Look for neurons that
receive input for the
flocculus (flocculus target
neurons, FTNs)
Caudal
Rat Brainstem Slice
AwayDay 2005
Slide No 28
Robotics
Does Algorithm Work in Real World?
• Tony Pipe, Chris Melhuish,
UWE Bristol
• Camera stabilisation
• How does algorithm compare
with control engineering
alternatives?
AwayDay 2005
Slide No 29
Multidisciplinary Approach
1. Framework: is cerebellar control of interest to robotics?
2. Problem: adaptive calibration of VOR
3. Approach: multidisciplinary
Modelling: plausible candidate algorithm
Electrophysiology: biological underpin
Robotics: real world application
AwayDay 2005
Slide No 30
AwayDay 2005
Slide No 31