Dean NMR 4

Question 1

How was the Marr-Albus Framework developed?

Answer 1

David Marr (1969) published a cerebellum model
James Albus (1971) published a theory of cerebellar function
The models aren't identical but generally considered sufficiently similar enough to be referred to as the Marr-Albus theory or framework

Question 2

What is the Marr-Albus Framework trying to do?

Answer 2

1. Explain why the cerebellum is important for motor control

It starts with trying to understand what motor control involves

Question 3

What technique is used to understand what motor control involves?

Answer 3

Task Analysis
Where complex information-processing problems need to be modelled at more than one level
The most abstract level is the computational level, which aims to simply say what problem the cerebellum is trying to solve

Question 4

What method is used in task analysis?

Answer 4

Starts with a simple task
Look initially straight ahead (O degrees) and target appears 10 degrees to the right
We look accurately at the target
We need to know how big a command to send to the muscles and eventually need to know how big a command to send to the muscle wherever the target is located
BUT how do we do this?

Question 5

What is the second level of the framework?

Answer 5

Marr’s algorithmic level:
stating that precise motor commands are learned using supervised learning
If it were a robot the algorithm could just be programmed in but can’t happen in humans and animals as we develop and change so require constant monitoring
This level is plausible because adults we make accurate movements ‘naturally’ in comparison to children who have to experiment and learn about their motor systems

Question 6

How can the correct value of a motor command be learnt?

Answer 6

A signal stating whether the movement is too big or too small would be helpful in the process
This can be derived from where the target is after you’ve made the movement
Have to make sure this is conveyed in order to initiate alteration (an error signal)

Question 7

What is the learning rule?

Answer 7

• If the movement is too small, make the command bigger
• If it’s too big, reduce the command
• If it’s accurate then do nothing

Question 8

How does the cerebellum carry out this error signal for learning?

Answer 8

A highly simplified explanation:
The parallel fibres signal the location of the target
When a parallel fibre fires, so does its target Purkinje cell (the motor command)
How much the Purkinje cell fires depends on the weight of the synapse between the parallel fibre and purkinje cell
If it is wrong then the weight is adjusted by the climbing fibre error signal

Question 9

What happens after enough learning trials in task analysis?

Answer 9

Will eventually learn how big a command to send to the muscles wherever the target is located
This is also true for a robot which demonstrates the power of the computational level

Question 10

What can this method also be referred to as?

Answer 10

The ‘function approximation’ method

This method learns to approximate the function that relates the motor command to retinal position

Question 11

What is the first reason that this learning rule model is plausible?

Answer 11

It explains in a general way why there might be so many granule cells
A lot of factors can affect the size needed for an accurate motor command, which is why there is the need for constant practice in music and sport
A lot of information is needed to get the movement just right

Question 12

What is the second reason for plausibility?

Answer 12

It accounts for why the climbing fibre input is so unusual
With the low frequency (1 spike/second) you don’t want the error signal to drive the movement directly
The CF input is very reliable which is an important trait in an error signal. So whenever a climbing fibre fires the PC also fires and the whole dendritic tree is affected

Question 13

How is this simplified?

Answer 13

The role of granule cells and ‘expansion recoding’ requires more detail
And the learning rule is slightly more complicated: if the PF signal is positively correlated with the CF signal then decrease the weight but if the PF signal is negatively correlated with the CF signal then increase the weight

Question 14

What is the learning rule sometimes called?

Answer 14

The decorrelation learning rule because learning ceases when there is no correlation between any PF input and the CF signal

Question 15

How does the learning rule/task analysis framework apply to eye blink conditioning?

Answer 15

In eye blink conditioning the initial movement is too small so the learning is always in one direction
The error signal (US) therefore needs to increase the movement size
Because of the circuitry this means the US should teach the Purkinje cells to fire LESS

Question 16

What happens in the cerebellum before conditioning?

Answer 16

The CS tone produces no change in cerebellar nuclei firing rates e.g. no conditional response, and no change in PC firing
This is possible because the excitatory (granule cells/parallel fibres) and inhibitory (stellate/basket cells inhibiting the PCs) inputs to the PC cancel each other out?

Question 17

What happens in the cerebellum during conditioning?

Answer 17

The error signal (US) reduces the weight of active parallel fibre synapses through LTD
If the inhibitory pathway is unaffected by climbing fibre input then the synapse is made progressively smaller until it doesn’t do anything at all
The net input into the PC is now minus and inhibits the cerebellar nuclei less and effectively takes the foot off the break
The CS tone now produces an increase in cerebellar nuclei firing rate
This makes functional sense

Question 18

What is the predicted site of plasticity?

Answer 18

The synapses between granule cells and purkinje cells

Question 19

What implications can be made about the cerebellar chip?

Answer 19

That the same basic circuit is used for eye blink conditioning as for movements in general
Therefore a cerebellar model used to explain the role in motor control should also apply to classical conditioning