is a schematic of the associative nature of the proposed plasticity.
James Albus had independently hypothesised that such plasticity
may take the form o f a long-term depression.
model has guided the majority of subsequent work on the role of the cerebellum in learning.
The model makes a number of predictions that are testable. The first of these is that the parallel fibre-Purkinje cell synapse should be plastic. Marr suggested that this was the only synapse within the cerebellar cortex that is capable of changes in efficacy. It now seems unlikely that only parallel fibre-Purkinje cell synapses in the cerebellum are plastic given that most synapses in the brain that have been investigated show some degree of plasticity. However, it could well be the case that the most important changes that occur during learning are at these synapses, and the Marr model predicts that interference with this plasticity should prevent motor memory formation (this will receive further consideration in the discussion).
Another correction to the Marr model was made only a few years after its conception, relating to the direction of the plasticity at parallel fibre-Purkinje cell synapses. Man- described the necessary alterations at these synapses as a potentiation of their
efficacy, akin to both Hebb’s original description of synaptic plasticity and long-term potentiation (LTP), the first real synaptic plasticity to be discovered in the brain (Bliss and Lomo, 1973). James Albus developed a model of cerebellar learning that was different in several ways from that of Man, not least in its description of the necessary plasticity (Albus, 1971).
Albus believed that potentiation of many parallel fibre-Purkinje cell synapses would lead to network instability. Furthermore, he realised that the output of the cerebellum as a whole, and therefore the expression of any learning in the cerebellar cortex that may occur, does not issue directly from the cortex. Rather, the output is relayed via one of several nuclei, either the deep nuclei in the cerebellum or the vestibular nuclei in the brainstem, that provide excitatory drive to other parts of the nervous system. The Purkinje cells provide the sole output of the cerebellar cortex, but they utilise the neurotransmitter G ABA (Obata et al., 1970) (Ito et al., 1970) and are therefore likely to be inhibitory in their action upon these nuclei. Any modifications in the cortex could only be expressed if they reduced this GABAergic action and thereby disinhibit the nuclei. Such a reduction would require a down-regulation in the parallel fibre
excitation of the Purkinje cell. Albus therefore suggested that the plasticity at parallel fibre-Purkinje cell synapses must take the form o f a long-term depression (LTD).
What are the predictions of the Marr/Albus model?
Masao Ito first demonstrated that these synapses are capable of LTD in the intact cerebellum (Ito, 1982a). His experiment consisted of conjunctively stimulating beams of parallel fibres and a climbing fibre. The result was a 20-50% reduction in the size of the excitatory post-synaptic potentials (epsps) elicited in the Purkinje cell by subsequent stimulation of the same parallel fibre beam alone. Such change was later shown to be input specific, because other sets of parallel fibres that were not
stimulated conjunctively with the climbing fibre elicited unaltered epsps (Ekerot and Kano, 1985), and there was also found to be no corresponding change in the climbing fibre elicited response. This depression was shown to be relatively long-term in that it lasts for over an hour.
For LTD to occur, however, it seemed that the stimulation frequency of the two different types of fibres had to be similar and in phase, conditions that were unlikely in vivo. Although the stimulation frequencies used initially were physiologically unrealistic, Ito’s finding was extremely important because it complied, at least loosely, with the predictions of the model. The result has since been replicated using different stimulation parameters (see (Linden and Connor, 1995) and (Daniel et ah,
1998)). Numerous experiments have been conducted in vitro and in vivo to assess the actual mechanism of cerebellar LTD. Further discussion of these experiments can be found in the final experimental chapter and discussion of this thesis.
If there was a way of recording the activity of Purkinje cells during learning, and of dissociating parallel fibre-initiated responses from climbing fibre input, then we would expect to find particular changes in the frequency of Purkinje cell responses to these two inputs. The Marr/Albus model makes certain predictions about how these two inputs would change relative to each other during learning. The first is that we might expect the climbing fibre activity, or instructional input, to increase in frequency during the learning of novel movement and decrease back to its original frequency after learning. The model assumes that the climbing fibre participates in the
learning o f a new movement but does not contribute to the m ovem ent itself. The second prediction is that the response o f the same Purkinje cell to parallel fibre input should be reduced in frequency as the task is learned and stay reduced during
subsequent performance o f the learned movement due to LTD. These predictions have in fact been tested in living animals using complex and simple spikes as
electrophysiological markers o f climbing fibre and parallel fibre input to Purkinje cells.
V
i20 ms
50 ms
I
1 mV
Figure 1.09: Electrophysiological markers - The above traces are different timescales o f an actual Purkinje cell recording in a ferret cerebellum. The asterisk indicates a complex spike, which is an electrophysiological marker o f climbing fibre input. The other action potentials are simple spikes. These spikes are considered markers o f parallel fibre input (Adapted from (Yeo and Hesslow, 1998)).
A complex spike, consisting o f a large initial depolarisation followed by a few smaller ones, is induced in a Purkinje cell by the climbing fibre due to its multiple contacts. The single synapse between each parallel fibre and the Purkinje cell can at most initiate a single simple spike post-synaptically in summation with many other parallel fibre inputs. Peter Gilbert and Thomas Thach (Gilbert and Thach, 1977) showed that a certain proportion o f extracellularly recorded Purkinje cells exhibited exactly the predicted alterations in responses during a motor learning task in monkeys. The monkeys were initially trained on a task in which they had to return a displaced
joystick handle to its initial position. In so doing they learned to compensate for a load placed alternatively on flexor and extensor muscles. Alteration of the load caused the monkeys to modify their compensatory movements in moving the joystick back to its central position. Learning to accomplish this was accompanied by transient increases in complex spikes that reduced as the movement was learned and subsequent long- lasting reductions in simple spikes.
A similar study using visually guided movement adaptation found changes in both simple and complex spike activity but no correlation between the two (Ojakangas and Ebner, 1992). Recordings in the flocculus of monkeys as they learn smooth pursuit eye movements concur in that uncorrelated changes are found during the learning (Kahlon and Lisberger, 1996). Both of these latter findings imply a role for the cerebellum in learning a new movement but do not adhere to the predictions of the Marr/Albus model because simple and complex spikes did not change in the same Purkinje cells.
A suppression of simple spikes has been observed in some eyeblink-related Purkinje cells of conditioned ferrets (Hesslow and Ivarsson, 1994). This finding bolsters the notion that parallel fibre-Purkinje cell synapses are depressed by learning (Albus,
1971) (Ito, 1982a). It is therefore consistent with a prediction of the model. All four of these electrophysiological studies vary in the degree to which they provide evidence for the validity of the Marr/Albus model, but they all contribute at least correlatory evidence for cerebellar cortical involvement in motor learning. As I have said previously, however, evidence that suggests a causal connection between neural alteration and behavioural changes should be considered over and above
straightforward correlations between the two. Presently there is no evidence to prove that memories are formed in the cerebellar cortex.
The deep nuclei may participate in learning
The neuroanatomy of the deep nuclei is less obviously specialised than that of the overlying cerebellar cortex. The nuclei also contain a much smaller number of cells. Models of their potential role in learning were not therefore immediately forthcoming. Only recently have theoretical suggestions as to how the nuclei may participate in the
process been presented (Mauk, 1997). These models have speculated as to the important cell interactions that may occur but have been unable to identify potential structural adaptations to information encoding akin to those found in the cortex. The a priori case for the deep nuclei as a site of motor memory storage is therefore not as compelling as for the cortex. Nevertheless, the empirical evidence demonstrating their central role in motor learning is impressive.
Anatomical studies have revealed limited but significant collateral inputs to the deep nuclei from both climbing fibres and mossy fibres. The inferior olive contributes input to the deep nuclei (Courville et al., 1977) (Dietrichs and Walberg, 1986), as do the pontine nuclei (Brodai et al., 1986) and reticular nuclei (Dietrichs and Walberg,
1979). There is therefore a potential parallel convergence o f contextual input and instructional input on the deep nuclei in motor learning just as Marr suggested for the cortex (Marr, 1969). It may be speculated that an independent parallel associative learning occurs in the deep nuclei between these two inputs. However, it has yet to be revealed that collaterals of fibres projecting from olivary and pontine origins converge on the same site within the nuclei. It is known that the inferior olive provides parallel input to the nuclei and those Purkinje cells that synapse at the same nuclear site (Ruigrok and Voogd, 2000). This finding presents several possibilities as to where associations could be formed. As well as the possibility that parallel learning occurs, it may be the case that stimulus input modified by the cortex becomes associated in the nuclei with either of the two collateral inputs, although it must be conceded, at least in the case of the pontine nuclei, that the collateral input is limited in size.
One particular model, originating from Michael Mauk and his associates, posits that at least some forms of motor learning result from changes within the deep nuclei. This model focuses on classical conditioning, although it aims to provide a general model of motor learning and considers adaptation of the VOR to be subserved in a similar manner. The CS is here held to be a component of the learning context, along with numerous other elements of the environment, body position and autonomic state. In the model CS input is delivered by mossy fibre collaterals and becomes associated with the US in the deep nuclei (Mauk and Donegan, 1997). The US input could be delivered by a variety of routes to the nuclei.
This proposal has been tested only in a simulated situation (Medina and Mauk, 1999). The principal aim of these tests was to evaluate the stability o f various possible associations that could be forged at mossy fibre-deep nuclear synapses. Changes at this synapse could potentially be governed by co-activity o f the climbing fibre collaterals to the same nuclear cells, the Purkinje cell input from the cortex, or in a purely Hebbian fashion by simply associating nuclear cell activity with mossy fibre input. It is important to consider not only how information might be encoded within the system by each of these mechanisms but also whether this information could be retained despite the continued activity of the system. Simulating each possibility on a computer demonstrates that synapses will, in all probability, slip back to their former state or be shifted into a novel state by the continued activity of the system. The only one of the cases considered in which this slippage will not occur is if changes in the weight of mossy fibre-nuclear cell synapses rely upon Purkinje cell input. Mauk’s model therefore suggests that US information is relayed indirectly via Purkinje cells from the cortex to the deep nuclei.
The inhibitory nature of the Purkinje cell-deep nuclear synapse has previously been regarded as puzzling but important. Mauk’s model seems to suggest that it is the inhibitory action of Purkinje cells that endows cerebellar-mediated learning with stability and may thereby solve this puzzle to some extent. A further feature of the model is that plasticity at mossy fibre-deep nuclear synapses is posited as bi directional. If the Purkinje cell input to the deep nuclei is co-active with the mossy fibre input then, within the model, a depression of the mossy fibre-deep nuclear synapse occurs. If the two inputs are not co-active then a potentiation occurs. Such bi directionality is considered to provide a means of accentuating the difference between associative and non-associative input (see fig. 1.10 - Bi-directional plasticity). These interesting proposals have yet to meet with universal acceptance. It is neither clear how an association between an inhibitory input and an excitatory input can be formed nor why it should be less susceptible to slippage after learning.