Corticocerebellar loops and volitional saccades: evidence for the role of the cerebellum in flexible behavioural control

(1)

Corticocerebellar loops and volitional saccades:

evidence for the role of the cerebellum in flexible

behavioural control.

Myrte Vos

July 2017

(2)

A thesis submitted in partial fulfillment of the MSc degree in Brain and Cognitive Sciences (University of Amsterdam). Supervised by Eric Avila Orozco, M.D., PhD (formerly of the De

Zeeuw group at the Netherlands Institute for Neuroscience) and Dr Harm Krugers.

Abstract

Though heretofore understood as a coordinator and optimiser of motor function, recent anatomical and functional evidence suggests that the cerebellum also forms closed loops with non-motor areas of the cerebral cortex, and therefore also plays potentially many roles in cog-nitive functions such as memory and goal-directed behaviour. The homogeneous architecture of the cerebellum implies that it processes motor and non-motor information in similar ways, though how exactly remains a topic of debate and ongoing research. In this thesis, experi-mental work seeking to clarify the possible ‘cognitive’ contributions of the cerebellum by way of a saccadic eye movement paradigm is discussed, and a review of corticocerebellar loops in the oculomotor system is given; finally, the accumulated evidence is related to the proposed Control Theory of cerebellar function.

(3)

1 Introduction

In this thesis, the phenomenon of interest is something we do thousands of times a day, automat-ically and unthinkingly: saccades, i.e. eye movements that are fast enough to allow us to focus and refocus on continuously changing visual targets without experiencing blurry vision.

Figure 1: Eye-movement tracking traces of saccades

The apparent banality of saccades belies their usefulness in cognitive neuroscience research: even something as simple as a split-second, subconscious decision about when to look where draws on neural networks that span the whole brain, from its stem to the prefrontal cortex, and may therefore reveal how similar networks subserve much more complicated functions (Hutton, 2008). Because the neural mechanisms underlying saccades are the best understood of any movement system (Robinson and Fuchs, 2001), very similar across vertebrates, and can be studied directly in animals using techniques such as single-neuron recording, saccade-based experimental paradigms are one of the few research methods available to us with which we can gain insight into the working of an awake, behaving brain at the cellular level.

Using the neural systems for saccade generation as a proof of concept, one can then explore the breadth of function of structures such as the cerebellum, which was believed to be involved primar-ily in the coordination and control of movement: particularly the acquisition and automatisation of motor skills (Marr and Thach, 1991; Albus, 1971; Ramnani, 2006). More recently, however, the cerebellum has been implicated in a range of non-motor functions (cf. Strick, Dum and Fiez, 2009), which raises the challenge of explaining its overall role in the brain in unifying terms - especially given its extraordinary homogeneous cytoarchitecture. The inputs and outputs of the cerebellum are now understood to form part of closed loops with other deep brain and cortical structures, which has led to the formulation of a cerebellar Control Theory (Ramnani, 2006; Ito, 2008; Imam-izu and Kawato, 2009). This theory, in which the cerebellum’s role in a closed corticocerebellar loop is to create, store, and iteratively refine internal models of the processes subserved by that loop, was originally developed with motor planning in mind, but may account for non-motor

(5)

pro-cesses equally well. Unlike motor function, however, the exact nature of cerebellar contribution to non-motor function are not yet well-understood. This is where saccades come in. As a motor process for which the involved sites in the cerebellum are amply documented (cf. Sun, Barash and Thier, 2015), volitional saccades can help elucidate any contributions of the cerebellum to the non-motor aspects of their execution. Anti-saccades, which are made in the opposite direction of a given target and therefore require a conscious, planned inversion of a normal (‘pro’) saccade, are particularly interesting: they have been shown to rely on area 46 of the prefrontal cortex, which has in turn been shown to be connected with Crus II, an area in the lateral cerebellum (Kelly and Strick, 2003). Does this mean that the cerebellum is not only directly involved in executing the saccade movement correctly, but also in preparing it in accordance with top-down instructions?

1.1 Scope and aim of this thesis

The above offers a sparse illustration of how the study of one small aspect of one area’s neural activity (saccade-related activity in the cerebellum) can feed into a broader understanding of neural systems (closed corticocerebellar loops in the oculomotor system), which in turn can inform models of cognition (Control Theory). This thesis is an attempt to examine that one small aspect (an experiment done by my supervisor), in the context of a review of those neural systems. It was written under the auspices of Dr Eric Avila Orozco, whose electrophysiological recordings of cerebellar Purkinje cells in macaques using a pro- and anti-saccade experiment paradigm were done to clarify the role of the cerebellum in saccade generation, and ultimately in cognition (Orozco, 2015). His work and results are discussed in section 2, in relation to my own partial, qualitative analysis of a batch of unpublished data he collected from the same monkeys using a modified (and cognitively more demanding) version of his experiment. Section 3 aims to give a functional and basic anatomical description of the corticocerebellar loops, and their separate components, that are involved in the production of volitional saccades. In section 4, I will attempt to relate these loops to cerebellar control theory, and discuss how recent saccade research elucidates the role of the cerebellum in encoding cognitive information.

1.2 Using volitional saccades to study flexible behavioural control

In primates, sharp vision is enabled through a small area in the center of the retina with a high concentration of cone cells. This area, the fovea, at most only captures 2 degrees of the visual field; the eyes must therefore be able to rapidly reorient themselves to align the fovea with a target of interest. These rapid adjustments are called saccadic eye movements. Figure 1 shows (by tracing the path of a person’s gaze as they view an image of a face, captured using eye-movement tracking) how the eyes do not inspect an object ‘holistically’, with a steady gaze, but by darting between visually salient points in an almost connect-the-dots fashion. Saccades are made at an average rate of 3/s, at a speed of up to 700o/s (up to 1000 in primates), with a duration of ¡50ms, with

(6)

little to no conscious effort; because of their speed, saccades are executed without online visual or proprioceptive feedback. They are therefore driven by constant monitoring and instruction from the oculomotor system (Guthrie, Porter, Sparks et al., 1983; Robinson and Fuchs, 2001). That system enters a period of ‘dead time’ ( 100ms) just before saccade onset, in which the saccade is already programmed and can no longer be adjusted to a sudden change in target (Findlay, 1997). There are many different types of saccades, all generated by the same basic neural circuitry (composed principally of the brainstem, the cerebellum, and the superior colliculus) and modulated by inputs from areas such as the basal ganglia, thalamus, anterior cingulate cortex and prefrontal cortex, depending on circumstances. Whether a saccade is made to scan the environment, follow a target, read a text or react to a sound, it generally comes in two types: reactive and volitional. Reactive saccades are nearly reflexive; they’re made in response to an appearing target, have a latency of about 150ms, and can under some circumstances be as fast as 100ms (Fischer et al., 1993). Volitional saccades, i.e. saccades that are internally generated and deliberately directed, are much slower. The superior colliculus is the main instigator of saccadic eye movements: it takes around 40ms for a visual stimulus to be transmitted from the retina to the superior colliculus, and another 20ms for that stimulus to result in a saccade. However, the observed latency from stimulus to a voluntary saccade is usually around 200ms, because other cognitive processes, such as decision-making, memory, and the suppression of distractors, contribute to but also thereby delay the generation of a saccade (Glimcher, 2003.

One type of volitional saccade that is often used to study cognitive control of movement is the antisaccade, in which a saccade must be made to the opposite direction of a given target. Unlike the prosaccade, the antisaccade requires top-down behavioural control; in order to perform it correctly, one must suppress the automatic response to look at the target (prosaccade) and then transform the location of the stimulus into a voluntary motor command to look away from the target (antisaccade) (Munoz and Everling, 2004).

Models developed to explain varying saccade latencies posit that in order for a saccade to be initiated, neural activity must increase from a baseline until it exceeds a threshold; many potential saccade targets may compete for attention at any time, and so the most salient target, for which the neural activity exhibits the fastest increase in firing rate, wins the ‘race’ (Noorani and Carpenter, 2013).

In an antisaccade task, neural signals for both the pro- and the antisaccade must race to threshold: the antisaccade is at a disadvantage, because the spatial inversion of the prosaccade must first be computed (Zhang and Barash, 2000), and must therefore rely on a third ‘suppression’ signal to stop the prosaccade in order to win (see Figure 3). As such, this task is relatively vulnerable to error, resulting in lower accuracy, longer reaction times and lower velocities compared to prosaccades (Bell, Everling and Munoz, 2000).

(7)

Figure 2: Diagram of a monkey performing a prosaccade (left) and an antisaccade (right). Figure reproduced from Orozco, 2015.

Figure 3: Proposed model for antisaccades, reproduced from Noorani and Carpenter, 2013. The common 60ms delay corresponds to visuomotor processing of the saccade target; the 50ms delay of the antisaccade corresponds to the spatial inversion.

processes: by comparing the neural activity of an area of interest during the performance of pro-and antisaccades, it is possible to tease out how that area contributes to the flexible behavioural control of the saccade generation process (Everling, Dorris, Klein and Munoz, 1999, Everling and Munoz, 2000). This is often done in primates by tracking eye movement during saccade tasks, while recording the electrophysiological activity in single neurons with electrodes; or by using those electrodes to stimulate neurons, in order to study the resultant behaviour (see Table 1 in the appendix for an inevitably incomplete overview of such studies using pro- and antisaccade tasks).

1.3 The cerebellum

Underneath the large, oblong cauliflower that is the human brain (the cerebrum) sits, like a shelf mushroom on the brainstem, a smaller structure affectionately named the cerebellum, or ‘little

(8)

brain’. In fact, due to the enormous amount of tiny granule cells packed inside it, the cerebellum contains 80% of the brain’s neurons, at only 10% of its volume. This economy of space is made possible by the cerebellum’s highly homogeneous cytoarchitecture (see Figure 5), which, like the motor cortex, is organised somatotopically.

Figure 4: Gross anatomical division of the cerebellum. Right: sagittal slice with cerebellar nomen-clature as defined by Larsell (1952). Reproduced from Orozco, 2015 (p.18).

The cerebellar cortex has three layers: the topmost molecular layer, the Purkinje cell layer and the granular layer. Most inputs to the cerebellum come from mossy fibres from the pontine nuclei (sensory-motor signals), but also from the brainstem (vestibular and reticular signals) and the spinal cord. Mossy fibres synapse with the granule cells (which are inhibited by Golgi cells). Together, these trisynaptic interfaces are called glomeruli (from the Latin: ‘little ball of yarn’). One mossy fibre excites between 400-600 granule cells. The parallel fibres (axons from the granule cells), extend upwards and spread laterally across the molecular layer. A single parallel fibre con-nects with between 300-450 Purkinje cells (PC), with as many as 180,000 individual synapses, and with a single PC receiving input from up to 200,000 parallel fibres. The other source of input to the PC comes from the climbing fibres (CF), carrying signals from the inferior olive: each PC receives input from only one CF, but a single CF contacts up to 10 PC (De Zeeuw et al., 2011). PCs output to the deep cerebellar nuclei. PC activity is inhibitory and more or less continuous: they produce two types of electrical signals, called simple spikes and complex spikes. Simple spikes are ˜1ms long and are regularly produced at a rate of 40-90 Hz, with a maximum of ˜250 Hz (De Zeeuw et al., 2011). Complex spikes have a regular firing rate of 1-2 Hz, with a maximum of ˜12 Hz, and are caused by CF activation. The PCs form the sole output from the cerebellum, so the patterning of simple and complex spikes in the PC encodes all information that is sent out to the rest of the brain.

While the cerebellum was initially considered to be ’just’ an overseer of motor function, sus-picions of cerebellar involvement in a range of nonmotor functions were first raised in the 1980s (Leiner, Leiner and Dow, 1986): there is now growing evidence of cerebellar activity relating to

(9)

Figure 5: Cytoarchitecture of the cerebellum. 1 = Purkinje cell, 2 = granule cell, 3 = stellate cell, 4 = basket cell, 5 = Golgi body, A = climbing fibre, B = mossy fibre, C = parallel fibre. Diagram by Wikipedia user Iamozy (retrieved from the article on glomeruli).

attention, emotion, memory, learning, and language (e.g. Baillieux, De Smet, Paquier, De Deyn and Mari¨en, 2008; Strick et al., 2009; Balsters, Whelan, Robertson and Ramnani, 2012; Buck-ner, 2013), and besides cognitive symptoms observed after cerebellar lesions (Schmahmann and Sherman, 1998), abnormalities of the cerebellum are now being associated with developmental cog-nitive disorders such dyslexia (Nicolson and Fawcett, 2005), schizophrenia (Andreasen and Pierson, 2008), and autism spectrum disorder (Rogers et al., 2013). These findings compel a serious over-haul of ’cerebellar doctrine’ (Galliano and De Zeeuw, 2014), of which many other fields of research, from cognitive science and neuropsychiatry to AI, stand to gain a great deal.

1.3.1 Cerebellar control of saccades

That the cerebellum contributes to the control of eye movements is one of the earliest-known of its roles: Hoshino (1920) electrically stimulated the cerebellar vermis of a rabbit to elicit eye

(10)

movement, and a large volume of research has since ensured that its involvement in the oculomotor system is among the best-documented of the cerebellum’s activities, though the jury is still out on the exact mechanisms by which it does so. The consensus so far is that, rather than initialising saccades, the cerebellum presides over the accuracy of their execution. Lobules VI and VII, aptly named the oculomotor vermis (OMV), have been robustly identified as being heavily involved in saccade control (Thier, Dicke, Haas, Thielert and Catz, 2002). Single-neuron recording studies (performed by surgically inserting a thin probe into the brain of an awake, immobilised animal, to measure the electrical activity of purkinje cells) have established that neurons in the OMV generally exhibit a burst (i.e. a large increase in the simple spike firing rate) of activity around the eye movement. At least half of recorded neurons display a heightened activity that starts 20-40ms before saccade onset, is correlated with saccade duration, and persists for up to 150ms afterwards (Ohtsuka and Noda, 1990; Thier et al., 2002; Kojima, Soetedjo and Fuchs, 2010). Pauses (large decreases in the simple spike firing rate) are also observed, with most recorded neurons exhibiting a mix of bursts and pauses before and during the saccade.

Besides the OMV, saccade-related activity has been noticed in the lateral cerebellum as early as 1973 (Ron and Robinson, 1973), particularly Crus I and II - though that line of investigation has not been intensely pursued (more on that later). Mano, Ito and Shibutani, 1991 observed simple spike modulation in PCs in the lateral hemispheres starting as early as 100ms before saccade onset; much earlier than in the vermis. Lesions in the lateral cerebellum of monkeys caused delayed onsets and dysmetria of saccades (Ohki et al., 2009).

(11)

2 Experiments with electrophysiological recordings in the

cerebellum

The cerebellum’s aforementioned suspected role in cognitive processing (particularly those requir-ing precision timrequir-ing: e.g. Peterburs and Desmond, 2016; Teki, Grube, Kumar and Griffiths, 2011) has reinvigorated efforts to understand its behaviour on the cellular level. The smoking gun of cerebellar lesions and satellite photography of functional imaging are not enough to tease apart whether cerebellar areas like Crus I/II actually contribute to the cognitive preparation of a saccade, or whether their activity just reflects online control of eye movement (Glickstein, Strata and Voogd, 2009). Direct recording of the electrical activity of task-relevant Purkinje cells can hopefully help answer that question.

Viral tracing studies of the cerebellum’s connections with other brain areas give a hint as to where to look. Kelly and Strick, 2003 have found vermal lobules VII and IX and Crus I/II of the cerebellum to form part of a closed-loop system with the area 46 of the prefrontal cortex: its mossy fiber inputs relay prefrontal signals from the pontine nuclei, and its outputs from the cerebellar nuclei circle back to the prefrontal cortex via the thalamus. The topographical organisation of the loop is such that the same cerebellar regions that receive indirect input from area 46, also project to area 46 (Kelly and Strick, 2003; for a detailed description of this loop, see section 3.2).

The experiments in this section were conducted with the aim to compare the activity of Purk-inje cells in Crus I/II, which was hypothesised to be involved in deciding on and/or planning of internally generated saccades (Ron and Robinson, 1973, Ohki et al., 2009, Ashmore and Sommer, 2013), with that in lobules VIc and VIIa of OMV, whose direct control of the timing, amplitude and sometimes also direction of more reflexive saccades is a fairly robust finding (Kojima et al., 2010). The pro- and antisaccade task, as discussed earlier, is a good way of differentiating between the roles of these two cerebellar areas: whereas prosaccades can be considered a semi-reflexive response to an external stimulus, antisaccades require conscious control of an internally generated movement (Munoz and Everling, 2004). An additional plus is that manipulation of the dorsolateral prefrontal cortex (the area found to be reciprocally connected to the lateral cerebellum) affects the correct performance of antisaccades, but not prosaccades (Johnston and Everling, 2006; Johnston, Koval, Lomber and Everling, 2013).

The experiments (a classic pro- and antisaccade task, see section 2.1, and a variation with a slightly more complex cue for task instruction) were conducted by then-PhD candidate Eric Avila Orozco, M.D., at the Netherlands Institute for Neuroscience, in collaboration with Peter Holland, Moshe Godschalk, Chris van der Togt, Peter Thier, Maarten Frens, and Chris de Zeeuw. They hypothesised that: ”Purkinje cells in the oculomotor vermis should modulate predominantly at the onset of or during the saccade in both the prosaccade and antisaccade task and that the Purkinje cells in the lateral cerebellum should modulate relatively strongly in the decision or instruction process during the antisaccade task, long before the actual saccade takes place.” (Orozco, 2015, PhD dissertation, chapter 5.)

(12)

2.1 Pro- and antisaccade Purkinje cell simple spike activity in monkey

oculomotor vermis and lateral cerebellum

Note: this section is an abbreviated and summarised version of the full chapter 5 of Orozco, 2015. Inevitably, the two texts are therefore very similar in places. I have also elected to leave out certain technical details and aspects of the analysis and results that I considered peripheral to the thesis’ main narrative. Any resulting mistakes are thus mine and not Eric’s.

Methods and materials

Two adult male rhesus monkeys (macaca mulatta) were prepared for eye movement recording and awake extracellular single unit recordings in the cerebellum using surgical and electrophysiological techniques in a two-step procedure. Under general anesthesia, a titanium head holder was im-planted to painlessly immobilize the monkey’s head. Four months later, once the monkeys had mastered the behavioural task (see below), a custom-made 40mm chamber was implanted to gain access to the cerebellum with a 25 degree angle (Figure 6). The animals recovered for at least 21 days before training was resumed. All procedures complied with the NIH Guide for Care and Use of Laboratory Animals (National Institutes of Health, Bethesda, Maryland), and were approved by the institutional animal care and use committee of the Royal Netherlands Academy of Arts and Sciences.

Figure 6: Implanted cerebellar access chamber. Reproduced from Orozco, 2015.

The monkeys were trained to perform a randomised interleaving pro- and antisaccade task in 8 different directions (cardinal and diagonal), with amplitudes between 5 and 14 degrees. All record-ings were conducted with the monkey in complete darkness. During training and experiments, the monkeys were seated in a primate chair with their head restrained at 100 cm from a screen with a resolution of 1,024 x 768 pixels. Visual stimuli were presented by a CRT-Projector Marquee 9500 LC, with a refresh rate of 100 Hz. The monkeys had unrestricted binocular vision.

(13)

Hi-Speed Primate, SMI GmbH, Germany). Single-unit recordings were obtained using tungsten glass-coated electrodes (1-2 M, Alpha Omega Engineering, Nazareth, Israel) inserted through a 23-gauge guide tube, which was inserted only through the dura. A motorised microdriver (Alpha Omega Engineering, Nazareth, Israel) with a 1-mm spaced grid was used to introduce the electrode and map the recording sites with a maximum resolution of 0.25 mm.

A trial started when the monkey focused his eyes on a red or green fixation point at the center of the screen for a random time between 300-500 ms. Then, a red target appeared in one of the 8 different target locations. After 100ms the fixation point changed to gray and, depending on its previous color, the monkey had to perform either a prosaccade toward the target (red), or an antisaccade to the opposite direction with the same amplitude as the target (green) within 500ms. The monkeys received a reward (a sip of juice) if they performed a correct saccade within 6 degrees of the (anti-)target and maintained fixation for 100ms. Within a single trial block the 16 possible configurations (8 targets * 2 saccade types) were presented in a random order. Blocks were presented as long as the quality of the recording was sufficiently good.

Data analysis

All analyses were performed off-line using custom programs written in Matlab (The Mathworks, Natick, MA, USA). Analysis was done on the kinematic parameters of the saccades as well as Purkinje cell activity and modulation in the OMV and lateral cerebellum for both the pro- and antisaccade task. Saccade onset and offset were detected using an adaptive threshold and computed as described in Nystr¨om and Holmqvist, 2010. For each trial, noise was removed and trials with reaction times less than 80ms were excluded, as these were thought to represent anticipatory actions. Saccade reaction time, amplitude, peak velocity and duration were extracted. Neural firing rate was estimated using a continuous spike density function generated by convoluting the spike train with a Gaussian function of σ = 20ms width (Silverman, 1986). Only saccade-related neurons were included: a neuron was considered to be saccade-related when complex spike and/or simple spike firing during the first 100 ms of the eye movement was significantly different from the preceding baseline activity (Wilcoxon rank test; p <0.05) for at least one of the 8 directions. Population spike density function was computed by calculating the firing rate of every Purkinje cell during the length of the trial by convoluting a Gaussian function of σ = 20ms width and averaging all the individual spike density functions (Dash, Dicke and Thier, 2013).

The characteristics of the saccade-related activity, denoted as a burst or pause, or a combination of both, were identified with an algorithm described in Cocatre-Zilgien and Delcomyn, 19921_Trials

were separated into four groups of neurons based on site of recording (OMV or lateral cerebellum)

1_{Briefly, a histogram of interspike intervals was constructed, while the identification of a threshold value was} obtained by analysing the transition from a burst to a non-burst state. Each value was then compared with the threshold value to identify the boundary spikes representing the start and the end of a burst (Cocatre-Zilgien and Delcomyn, 1992).

(14)

Figure 7: Example simple spike ’rasters’ from the same OMV neuron (left = prosaccade, right = antisaccade). Blue and red lines (top) represent eye kinematics for all trials. The black dots (bottom) are simple spikes. The single, slightly tilted vertical blue line represents the ’go’ cue (monkey permitted to make a saccade); the red line represents simple spike firing rate averaged over all trials. y-axis = trial number, z-axis = firing rate, x-axis = time (seconds).

and instruction (prosaccade or antisaccade). Only correctly performed trials were analysed. Purkinje cell activity was determined for four different time windows, including 1) the baseline period, which spans the intertrial interval, i.e. 400ms before fixation until fixation; 2) the instruc-tion period, during which the monkey was fixating on the center point but already instructed by the color of this fixation point to generate either a pro- or antisaccade (this period ended 100ms before the target appeared); 3) the direction period, during which the monkey was still holding fixation and simultaneously observing a target in one of eight different directions for 100ms; and 4) saccade- execution period, during which the monkey was making the saccade (i.e. measured 25ms before saccade initiation to 250ms after saccade initiation). A neuron was considered to be task-instruction related when the firing in the instruction window was significantly different from the preceding baseline activity (Wilcoxon rank test; p <0.05). Statistical comparisons were performed using a paired t-test with Bonferroni correction or a two-tailed Student’s t-test.

Results

In accordance with earlier studies (e.g. Bell et al., 2000), antisaccades had longer reaction times, larger amplitudes, lower peak velocities, and overall longer durations than prosaccades. Also

(15)

expected was the correlation between larger saccade amplitudes and longer duration. There was no correlation between reaction time and any other kinematic parameters.

In OMV, 27 neurons were recorded. Baseline activity during pro- and antisaccade tasks was not significantly different, and only one neuron presented significantly higher activity during the task instruction window for antisaccades compared to prosaccades. (see Figure 8).

Simple spike activity during the direction instruction window (100ms; after task instruction, but before saccade initiation) of both pro- and antisaccade tasks was examined. Among the neurons recorded, 25.9% (n = 7/27) of the neurons presented significant differences. Although the polarity of these changes in responses varied, (11.1% (n = 3/27) of the neurons increased their firing rate, whereas 14.8% (n = 4/27) decreased) this suggest that some PCs in OMV may encode the direction of saccades. Finally, neurons were classified based on their general features (i.e., bursting and pausing behaviour) during the saccade execution window. A significant burst or pause onset was calculated to be 5 times the standard deviation of baseline firing rate. 48% showed bursting-only, 13% pausing-only, 13% burst-then-pausing, and 26% pause-then-bursting activity. Qualitatively, all these neurons showed a consistent classification during the two tasks, despite the facts that their firing rates changed during the actual saccades and that these could vary dependent on the task involved.

Overall, 47% of neurons (n = 13/27) presented significantly different responses between tasks; 69% of which (n = 9/13) showed significantly higher simple spike activity for the antisaccade compared to the prosaccade task. Firing rate peaks were also on average later during prosac-cades compared to antisacprosac-cades. In both tasks, the majority of neurons presented a large burst around/right after saccade onset.

Figure 8: Statistical summary of neural activity in OMV and lateral cerebellum in pro- and antisaccade trials. Reproduced from Orozco, 2015.

In Crus I/II, 31 neurons were recorded. 9.6% of the neurons (n = 3/31) showed significant differences in firing rate during the baseline period (all for the prosaccade task). During the in-struction window 35.4% (n = 11/31) of neurons showed significant differences between prosaccades and antisaccades, 72.7% (n = 8/11) of which had higher firing rates during the antisaccade task. During the periods of direction instruction 19.3% (n = 6/31) of neurons presented significant dif-ferences between the two tasks, most of which (83.3%, n = 5/6) also presented higher firing rates for antisaccades.

(16)

Figure 9: Simple spike firing rate distributions. Reproduced from Orozco, 2015.

Like the neurons in OMV, neurons in the lateral cerebellum had internally consistent fir-ing patterns. The distribution of pattern types was: 36% burstfir-ing-only, 42% pausfir-ing-only, 1% burst-then-pausing, and 21% pause-then-bursting activity. Overall, 45% of the neurons in Crus I/II presented significantly different responses between the two tasks (n = 14/31), half of which presented higher simple spike activity for the antisaccade task. In contrast to the neurons in the OMV, neurons in the lateral cerebellum showed more pausing activity around saccade onset and these pauses lasted longer. Generally, activity in this area began and ended several 100ms before and after saccade onset, respectively; and, as can be seen in Figure 9, was less densely concentrated around the saccade itself, showing more variability in spike rate throughout the task.

In conclusion, this study revealed that there is significant saccade-related activity in Purkinje cells in both the oculomotor vermis and Crus I/II of the lateral cerebellum, and moreover, that between 1/3 and half of these neurons modulate their activity depending on whether the saccade is pro or anti.

(17)

2.2 Qualitative analysis of a 4-colour variation on the pro- and

antisac-cade experiment

One intriguing outcome of the experiment described above was that 35% of neurons in the lateral cerebellum modulated their activity between pro and antisaccades during the instruction window of the task, while the monkey was fixating, before it had even received directional instructions (and could therefore begin to prepare an eye movement to any particular location). That means that these neurons were engaged in a part of the task that could not be related to the saccadic movement itself - only the processing of the first-order rule for task instruction (red = pro, green = anti), and the subsequent anticipation of having to suppress a prosaccade and invert it by 180 degrees. In short, cognitive, non-motor processing.

Figure 10: Experiment paradigm for a pro- and antisaccade task with 4 rather than 2 possible fixation point colours. Figure by Eric Avila Orozco.

To further investigate this promising finding, Eric did another experiment, the results of which have (to my knowledge) not been published. In this more complex variation of the pro- and anti-saccade task, the fixation point was presented in four colours instead of two: purple, blue, yellow and orange. The rules for which colours represent either task instruction were changed for each

(18)

recording session, requiring the monkey to learn them by trial-and-error and maintain them in working memory throughout the session.

Figure 11: Two examples of the processed data rasters. A single raster represents all the trials for one task type (pro or anti) for one neuron over one recording session. The red line is the simple spike firing rate of the recorded Purkinje cell. Unlike the rasters in Figure 7, these rasters are aligned with fixation cue onset (t = 0), meaning that -0.5 <t <0 is an intertrial interval in which the monkey is idle. The yellow box delineates the 400-600 ms before initiation of the saccade, a.k.a. the instruction period. The left raster is a prosaccade trial in the lateral cerebellum; the right is an antisaccade trial in OMV.

I did a qualitative analysis of the instruction window of this 4-colour paradigm. Because of considerable limitations in time and resources (crucially, I had no programming ability whatsoever and was also largely unfamiliar with the statistical and mathematical tools required to perform the analyses used by Eric and Peter Holland in Orozco, 2015), I did not run any further quantitative analyses. I examined each raster in the dataset and estimated the mean baseline firing rate and the lowest and highest firing rates in the analysis window (100-300ms after fixation cue onset), described any other noticeable phenomena or patterns, and based on that classified each raster as exhibiting one of five possible behavioural patterns: burst, pause, burst-then-pause, pause-then-burst, or none.

(19)

Figure 12: Example rasters of neurons in OMV and lateral cerebellum, exhibiting three common pattern types. Neurons were usually internally consistent across pro- and antisaccade trials.

The distribution of different pattern-types that I found for this dataset is given for each task type and recording location in Figure 13. ’None’ means there was no significant modulation during the instruction window, but that doesn’t mean it’s just noise: you tend to always see a peak/pause around 100ms, because the neuron responds to seeing the fixation point. Similarly, there is often a steep rise or fall in activity close to 500ms, when the saccade is imminent.

(20)

Figure 13: Results of the qualitative analysis of the firing pattern of Purkinje cells in OMV and lateral cerebellum during the instruction window of a 4-colour pro- and antisaccade task experiment.

A considerable proportion of neurons in both OMV and lateral cerebellum showed signific-ant modulation during the instruction window of both pro- and signific-antisaccade trials, compared to baseline activity before the fixation cue (t = 0). In lateral cerebellum, that proportion constituted the majority of neurons. Neurons in both OMV and lateral cerebellum showed a high level of consistency in proportions of pattern-types across the two tasks, suggesting that at least in these neurons, the content of the instruction doesn’t effect significant differences in activity - that, pre-sumably, happens further upstream, in the prefrontal cortex.

One surprising and unanticipated observation to emerge from this dataset were the patterns seen in some of the target calibration (”CalTarg”) rasters (see Figure 14). The CalTarg rasters were composed of trial blocks in which the monkey just made simple prosaccades over and over, without the task. The process was essentially the same as for a ’real’ prosaccade task: fixate, wait, saccade, reward. But the observed activity was often completely different; sometimes even inversed. This suggests that the mental context of the task matters, at least to some neurons. (CalTarg trials were unfortunately not recorded during the 2-colour pro- and antisaccade experiment, so these rasters cannot be compared to any equivalent data in the cognitively less demanding version of the experiment.)

(21)

Figure 14: Target calibration rasters

In conclusion, the experiments described in this section demonstrate significant modulation of activity during pro- and antisaccade tasks, including one requiring the dynamic memorisation and correct application of a cue rule, in Purkinje cells in both an established cerebellar oculomotor area, and a cerebellar area associated more with non-motor processing. Preliminary though it may be, the analysis of the 4-colour paradigm instruction window articularly suggests that neurons from both regions play a role in the non-motor processing of pro- and antisaccades, though what role exactly is not clear - error monitoring, information retrieval, memory activation or something else. A fascinating hint of an unforeseen factor influencing neural activity is given by the target calibration trials, suggesting that the modulation of these neurons is dependent on the mental context in which a task is performed - a level of discernment which was not expected from the ’floor manager’ of motor behaviour.

(22)

3 Corticocerebellar loops and volitional saccades

3.1 Involved brain areas

3.1.1 Superior colliculus

The superior colliculus (SC), also known as the tectum, is a layered structure in the vertebrate midbrain the layers of which can be roughly classified as ‘superficial’ and ‘deep.’ The superficial layers represent visual information, and contain a topographic map of visual space organized in retinotopic coordinates. The deep layers are involved in movements of the eyes, head, and trunk (Wurtz and Albano, 1980). The function of the SC is to direct behavioural responses toward specific points in egocentric space: the SC receives direct projections from the retina, and sends signals to the brainstem that lead to the generation of saccades (see Figure 15).

Figure 15: Diagram of SC structure, reproduced from Carpenter, 1999.

Saccade neurons in the deep layers are inactive during visual fixation and discharge a high-frequency burst when saccades are made. Many also display low-high-frequency preparatory activity when there is a high chance of stimulus appearance, and a high-frequency stimulus-related burst of action potentials (Munoz and Everling, 2004; Johnston and Everling, 2008). Projections from the SC activate burst neurons in the paramedian pontine reticular formation (pPRF) and the rostral interstitial nucleus of the medial longitudinal fasciculus (riMLF), in the brainstem, which discharge high-frequency bursts of action potentials just before and during saccades, and innervate the neurons in the oculomotor, abducens and trochlear nuclei; these nuclei ultimately direct the muscles that move the eye. The SC also projects to the omnipause neurons of the caudal PRF, which act as a gating mechanism on the saccade generating system(Johnston and Everling, 2008). The SC receives input from a variety of cortical and subcortical areas, chief among which the frontal eye field (FEF), the supplementary eye field (SEF), the dorsolateral prefrontal cortex (DLPFC), the anterior cingulate gyrus (ACC), the lateral intraparietal area (LIP), and the basal ganglia. Lesioning of the SC results in increased response latency, and decreased velocity and

(23)

accuracy of saccades; when both the SC and the FEF are lesioned, fixation on visual targets and the range, velocity and frequency of saccades are seriously impaired (Schiller, Sandell and Maunsell, 1987), if the saccade is not eliminated entirely (Hanes and Wurtz, 2001). In monkeys performing a pro- and antisaccade task, SC was found to show increased activity during the task instruction period and decreased stimulus- and saccade-related activity, suggesting that the suppression of an unwanted prosaccade during an antisaccade task occurs at least partially in the SC (Everling et al., 1999).

3.1.2 Cerebral cortex

Several areas in the frontal, parietal and temporal cortex have been found to exert top-down cognit-ive influence on eye movements: for the purposes of this thesis, the areas of particular importance are the frontal eye field (FEF), supplementary eye field (SEF), dorsolateral prefrontal cortex or prefrontal eye field (DLPFC/PFEF), parietal eye field or lateral intraparietal area (PEF/LIP), and anterior cingulate cortex (ACC).

Figure 16: Interconnections of cortical oculomotor regions and caudal ventrolateral prefrontal cortex of the macaque. Reproduced from Borra, Gerbella, Rozzi and Luppino, 2013.

Frontal eye field

Located in the precentral gyrus, also known as area 8. The frontal eye field disengages fixation, and directs intentional saccades to visible targets, to remembered target locations, or to the location where it is predicted that the target will reappear (Pierrot-Deseilligny, Rivaud, Gaymard, M¨uri and Vermersch, 1995). It plays an important role in modulating visual attention, and in the smooth pursuit of moving visual targets. Microstimulation of the FEF has given rise to the theory that the neuronal activity in a particular area represents a retinotopic goal with reference to an earlier eye position, which is converted by a downstream structure (probably the SC) into the vector of an eventual saccade (Dassonville, Schlag and Schlag-Rey, 1992, Hanes and Wurtz, 2001). FEF neurons discharge in response to visual stimuli, whether they result in a saccade or not, and also display prestimulus activity that has been shown to differ between pro- and antisaccades (Everling and Munoz, 2000). Bilateral damage to the FEF impairs the accuracy and latency of saccades,

(24)

and can also cause visual perception deficits; however, these symptoms are transient, and within weeks are presumably taken over by other elements of the oculomotor system (Lynch Tian 2006). FEF projects heavily to the SC and, to a lesser extent, other brainstem oculomotor structures, though FEF is not able to drive saccade generation without the SC. It also has reciprocal con-nections with the basal ganglia (Alexander, DeLong and Strick, 1986; Hikosaka, Takikawa and Kawagoe, 2000), with the nucleus of Darkschewitsch and the parvocellular red nucleus in the mesodiencephalic junction, and with precerebellar areas of the pons (Borra, Gerbella, Rozzi and Luppino, 2013).

Supplementary eye field

Located in the dorsomedial part of the frontal lobe, the SEF does not appear to be directly involved in motor control or initiation of saccades (though if stimulated it does evoke them), but is considered responsible for the control of sequences of visually guided saccades (Gaymard, Ploner, Rivaud, Vermersch and Pierrot-Deseilligny, 1998), and of saccades complicated by simultaneous head or body movement (Pierrot-Deseilligny et al., 1995). While the saccades controlled by the FEF are visually guided, those controlled by the SEF appear to be internally guided: single-neuron recordings in the SEF reveal a higher firing rate before and during antisaccades compared to prosaccades (Schlag-Rey, Amador, Sanchez and Schlag, 1997). Likewise, the FEF shows almost no presaccadic activity for spontaneous saccades, as opposed to the SEF, which can increase its firing rate as much as 500ms beforehand (Amador, Schlag-Rey and Schlag, 2004). Like the FEF, the SEF has its own projections to the oculomotor brainstem, basal ganglia, and the parvocellular red nucleus; to what extent these connections are reciprocal is not yet fully understood.

Dorsolateral prefrontal cortex and ventrolateral prefrontal cortex

The prefrontal cortex is the seat of executive function, which includes working memory, attention, decision-making, and impulse control. The prefrontal eye field, which is located in the dorsolateral PFC, is involved in the inhibition and prediction of saccades, and spatial memory. Neurons in the DLPFC respond selectively to either the location of a stimulus, or the direction of an impending saccade; they were also found to increase their activation in preparing for an antisaccade, in com-parison with prosaccades. This heightened activity is signalled directly to the SC, where it was thought to stimulate fixation neurons and thereby suppress a reflexive prosaccade (Everling and DeSouza, 2005; Johnston and Everling, 2006); however, more recent work challenges that interpret-ation (Johnston et al., 2013), suggesting instead that the DLPFC’s inputs to the SC are excitatory and represent attentional shifts (Kaping, Vinck, Hutchison, Everling and Womelsdorf, 2011). A review of the role of the prefrontal cortex in cognitive control concluded that the DLPFC is likely involved in ’...the rule-based selection of responses [...], linking short-term memory representations to goal-directed motor behaviour [...], [and] successful decision making by using payoff expectation to guide the conscious and deliberate goal-directed selection of task rules and appropriate actions.’

(25)

(Ridderinkhof, Van Den Wildenberg, Segalowitz and Carter, 2004, p.133)

Investigating the specific contributions of the DLPFC to an antisaccade task with a working memory component, Hussein, Johnston, Belbeck, Lomber and Everling, 2014 found evidence of functional specialisation of two regions in the DLPFC: inactivation of the inferior DLPFC sab-otaged the working memory antisaccade task, but not the control task, and inactivation of the adjacent region had no effect on either. Deactivating both regions impaired both antisaccade tasks, but not prosaccade tasks. Temporary disruption of the DLPFC or the FEF in humans dur-ing a pro- and antisaccade experiment has demonstrated that the DLPFC exerts executive control over antisaccades, while the FEF controls their visuo-motor programming (Cameron, Riddle and D’Esposito, 2015).

The ventrolateral prefrontal cortex (VLPFC) is a cognitive control area which has only been recently investigated in relation to eye movements. It is associated with reflexive reorienting, motor inhibition, and action updating (Levy and Wagner, 2011). Its caudal portion contains several areas (Brodmann’s areas 8r, 45A, 45B, caudal 46vc and caudal 12r) which interconnect with frontal oculomotor regions FEF, SEF, and LIP. Its descending projections to the SC, basal ganglia, oculomotor brainstem and pons closely resemble those of the FEF, making it very likely that it participates in cerebellar and basal ganglia oculomotor loops, though its activity during oculomotor tasks has not yet been studied directly (with the exception of areas 8r and caudal 46vc, whose role in controlling visually and memory-guided saccades has been demonstrated by Funahashi, Chafee and Goldman-Rakic, 1993) (Borra et al., 2013). In a study with frontal lobe damaged patients performing an oculomotor rule switching task and the standard antisaccade task, damage to VLPFC was found to be a significant predictor of errors on both tasks, and it was suggested that VLPFC mediates inhibitory control over stimulus-elicited eye movements, with an additional specialisation for the monitoring and updating of task rules (Hodgson et al., 2007).

Parietal areas (LIP/area 7)

The lateral intraparietal area, or parietal eye field, is located in the intraparietal sulcus and is involved in visuospatial integration, in visual pursuit, and in triggering reflexive saccades in ex-ploration of the visual field (Pierrot-Deseilligny et al., 1995; Gaymard et al., 1998). LIP is active in anticipation of saccades made to the remembered location of both visual and acoustic targets. Like the FEF, ablation of the LIP will lead to only modest and transient impairments in eye move-ments, but lesioning both results in much more serious and long-lasting impairmove-ments, suggesting a measure of redundancy between the two areas (Lynch and Tian, 2006). Microstimulation of neurons in the LIP will evoke saccades even in the absence of the FEF; in a pro- and antisac-cades experiment, they encode visual location but not saccade direction (Gottlieb and Goldberg, 1999), but Zhang and Barash, 2000 have found that during certain memory-antisaccade trials, a subset of LIPs’ ‘visual’ neurons will discharge in the motor direction, suggesting a sensorimotor transformation from cue to motor instruction that takes place within 50ms.

(26)

Anterior cingulate cortex

The posterior part of the anterior cingulate cortex has only recently been implicated in the cognitive control of saccadic eye movements: lesions of the ACC result in impairments that are very similar to those caused by lesions of the FEF, SEF, and DLPFC (Gaymard et al., 1998). The ACC is thought to play a role in the motivation and preparation of intentional saccades; moreover, it has been shown that during alternating blocks of pro- and antisaccade trials, ACC neurons show high task-selectivity right after a task switch, which gradually decays as the same task is repeated (Johnston, Levin, Koval and Everling, 2007). The authors therefore suggest that the ACC and DLPFC both provide top-down instructions to the SC, and that the ACC is more active when task demands increase (Johnston and Everling, 2008); this theory is corroborated by microstimulation of the ACC during pro- and antisaccade trials, which has demonstrated improved performance and reduced reaction times for antisaccades (Phillips, Johnston and Everling, 2011).

3.1.3 Mesodiencephalic junction

The mesodiencephalic junction (MDJ) is an area located at the top of the midbrain, incorporating a variety of motor-related structures such as the nucleus of Darkschewitsch (nDark), red nucleus (RN), nucleus interstitialis of Cajal (inCajal) and the nucleus of Bechterew (nBech) (de Zeeuw et al. 1998). Some of these areas, like the magnocellular red nucleus (mRN), project directly onto the spinal cord, whereas the parvocellular red nucleus (the relative size of which, compared to the mRN, is much larger in humans than in other vertebrates - making it suspect in questions regarding human cognition) and the nuclei of Darkschewitsch and Bechterew project to the inferior olive. The MDJ receives much of its input from the deep cerebellar nuclei (De Zeeuw and Ruigrok, 1994). Neurons in the pRN also receive input from the primary, frontal, and supplementary motor cortices; the lentiform nuclei (BG); and from areas in the prefrontal, temporal, and occipital cortex. The contributions of the MDJ to the oculomotor system are not well-understood: neurons ventral to the nDark, rostral to the inCajal and rostral to the oculomotor complex of the medial longitudinal fasciculus have been shown to contribute to vertical eye movements (Sato and Noda, 1991), and lesions in the MDJ are linked to vertical gaze palsy, but virtually no work exists to demonstrate its role in saccadic movement. However, because the pRN does not display any movement-related activity, and given its rich connectivity with cerebellar, striatal, and cortical areas, it is hypothesized that it is somehow involved with cognitive control of movement (Nioche, Cabanis and Habas, 2009).

3.1.4 Basal ganglia and thalamus

As with corticocerebellar loops, the current understanding of the basal ganglia is that they form largely segregated, topographically organised circuits with a myriad of cortical areas, underlying motor, oculomotor, executive and limbic functions (Alexander et al., 1986). The basal ganglia do not control these functions directly, but are instead thought to play a modulatory role (Watanabe

(27)

and Munoz, 2011). Like the cerebellum, the basal ganglia were believed to receive information from a broad range of cortical areas, in order to ‘funnel’ that information to the primary motor cortex in the form of motor commands. However, based on retrograde transneural tracing of five regions in the prefrontal cortex (of which two, areas 9l and 46d, have been shown by the same authors to form part of closed corticopontocerebellar loops) of cebus monkeys, Middleton and Strick, 2002 estimate that one-third of all BG output is directed at the PFC – at least as much as its output to the motor cortex. Moreover, they found that all the prefrontal areas they examined formed closed loops with the BG: e.g., dorsal area 46 projects via the caudate nucleus to the internal globus pallidus (GPi), which projects back to 46d both directly and indirectly, via parvocellular ventroanterior thalamus.

The caudate nucleus mediates oculomotor projections from the DLPFC and the frontal and supplementary eye fields, which generate saccade instructions, to the thalamus and intermediate layer of the superior colliculus. The BG control the frontal signals to the SC through sustained inhibitory input via the substantia nigra pars reticulata (SNr), which can be lifted with inhibitory input from the caudate nucleus to the SNr (Hikosaka et al., 2000; Phillips and Everling, 2012). It is thought that caudate neurons use contextual information contained within their preparatory activity, such as stimulus-reward associations, to bias the selection of saccades to be further pro-cessed by the SC. Furthermore, studies of the BG ‘motor loop’ (comprised of the putamen, globus pallidus, ventroanterior and ventrolateral thalamus and the medial prefrontal cortex; previously thought to be functionally and topographically distinct from the BG oculomotor loop) suggest its involvement in the monitoring, and possibly the execution, of goal-directed saccades (Phillips and Everling, 2012). The putamen is implicated in monitoring performance error, and reward-history based action selection (Phillips and Everling, 2012; Muranishi et al., 2011). Neurons in the ventrolateral (VL) and ventroanterior (VA) thalamic nuclei mediate projections from the globus pallidus and SNr to the supplementary eye fields (SEF), and neurons in the mediodorsal do so from the superior colliculus and the SNr to both the FEF and the SEF (Lynch and Tian, 2006); single-neuron recording studies have found increased firing modulation during antisaccades com-pared to prosaccades in neurons of the VA/VL thalamus (Kunimatsu and Tanaka, 2010), the basal ganglia (caudate nucleus: Ford and Everling, 2009; globus pallidus: Yoshida and Tanaka, 2008; SNr: Gore, Marino and Munoz, 2005), and the SEF (Schlag-Rey et al., 1997), further supporting the hypothesis that corticostriatal loops are at least partially responsible for the correct execution of volitional saccades.

3.1.5 Cerebellum

The areas of the cerebellum that have been found to be involved with the oculomotor system are the oculomotor vermis, Crus I and II (ansiform lobule), the floccular complex, the uvula, and the nodulus. Of these, the first two are relevant to this review. Oculomotor vermis is involved in the control of saccades and smooth-pursuit eye movements and is located in medial and posterior lob-ules VI and VII (Voogd, Schraa-Tam, van der Geest and De Zeeuw, 2012). It receives its input from

(28)

climbing fibers originating from the caudal medial accessory olive (cMAO), known as subnucleus-b in monkeys and as subnucleus-c in rats. The cMAO, in turn, receives its input from the interme-diate and deep layers of the contralateral SC (Kandel, Schwartz, Jessell, Siegelbaum, Hudspeth et al., 2000). In rats, the terminal field of this projection in subnucleus c contains two neuronal populations: one projecting to lobule VII, and one projecting to Crus II, located in the medial and the lateral portions of subnucleus c of the cMAO respectively (Akaike, 1989). Lobule VII’s mossy fibre afferents originate from the nucleus reticularis tegmenti pontis (NRTP), the pontine paramedian reticular formation (PPRF), the prepositus hypoglossi and the dorsolateral pontine nuclei (DLPN) in the brainstem (Voogd and Barmack, 2006). These regions receive afferents from the SC and other subcortical oculomotor areas. Lobule VII’s efferents project to the caudal pole of the fastigial nucleus (cFN). The cFN also receives collateral projections from the climbing fibres terminating in lobule VII, and is reciprocally connected with the cMAO (Noda, Sugita and Ikeda, 1990). Purkinje cells in oculomotor vermis and cFN discharge spike bursts mainly during saccadic movement, regardless of type (Kojima et al., 2010) but with some sensitivity to direction (Oht-suka and Noda, 1990). Inactivation or lesioning of either of these areas will result in hypometric and more variable saccades: in cFN, the timing and firing rate of bursts appear to be related to the speed of the saccadic movement, which enables greater precision, and the timing of pauses in OMV Purkinje cells is hypothesized to trigger the bursts in the cFN (Robinson and Fuchs, 2001).

The lateral hemispheres of the cerebellum, where Crus II is located, are much larger in humans than in monkeys, having evolved towards rapid expansion together with the cerebral cortex from which they receive their input (Leiner, Leiner and Dow, 1991). Crus II is involved with saccadic eye movements and smooth pursuit, in particular when such movements require shifts in attention, memorizing target locations or saccade sequences, or learning a rule (Nitschke et al., 2004). It is innervated by mossy fibres from the prepositus hypoglossi and the medial and dorsolateral pontine nuclei; and by climbing fibres from the principal and rostral MAO (Voogd et al., 2012). The input to Crus II originates from a broad range of cortical areas, including known eye fields such as FEF, SEF and LIP, but also areas 5 and 7 in the parietal cortex and nonmotor areas such as DLPFC (Brodal, 1983; Kelly and Strick, 2003; Habas, Guillevin and Abanou, 2010). Visuomotor Purkinje neurons in the lateral hemispheres project to the cDN and the posterior interposed nucleus. These nuclei, in turn, project to the ventrolateral thalamus (VL), which connects to a variety of cortical terminals such as premotor and primary motor cortex, FEF, and LIP. The cDN also projects to neurons in the parvocellular red nucleus (in the MDJ). These neurons project to the inferior olive, which in turn projects climbing fibers back to the contralateral cerebellum, thus forming a closed loop (Kandel et al., 2000). Finally, cDN has a direct connection to the SC (May, Hartwich-Young, Nelson, Sparks and Porter, 1990). Neuronal recordings in the dentate nucleus during saccade tasks have demonstrated its involvement in ‘self-timing’ voluntary behaviour (Ashmore and Sommer, 2013), perhaps in tandem with the thalamus (Tanaka, 2007), FEF (Li and Lisberger, 2011) and LIP (Janssen and Shadlen, 2005), which have also been found to encode temporal aspects

(29)

of movement or sensation.

3.2 Corticocerebellar loops

3.2.1 Corticopontocerebellar loops

The enormous interconnectivity of the cerebellum with the cerebral cortex has long been well-established. It was assumed that the role of the cerebellum was to be on the receiving end of all these connections, transforming the information to plan and coordinate movement and projecting, via the ventralateral nucleus of the thalamus, to only one cortical area: the primary motor cortex (Strick et al., 2009). Transneural tracing techniques using neurotropic viruses (Kelly and Strick, 2003, 2000) have made it possible to trace multisynaptic networks, and thereby identify cerebel-lar efferents passing through functionally distinct subdivisions of the ventrolateral thalamus and fanning out to many cortical areas besides the primary motor cortex. We thereby know now that the cerebellum consists of multiple, anatomically distinct modules2 (Ramnani, 2006) that each form part of a closed loop to and from a specific cortical area: projections run from a particular cortical area via the pontine nuclei to the cerebellum, which in turn projects back to that same cortical area via the deep cerebellar nuclei and the thalamus (Middleton and Strick, 1998). The same principle appears to hold for cerebellar loops involving the olive and the midbrain (De Zeeuw et al., 1998; Voogd et al., 2012). Middleton and Strick, 2001 used the herpes simplex virus to trace, in retrograde, the projections from the cerebellum via the ventrolateral thalamus to the prefrontal cortex. Cerebellar outputs were seen in areas 9 and 46 (dorsolateral prefrontal cortex), and the dentate regions from which these connections originate were shown to be separate and distinct from those that project to the motor cortex (see Figure 17).

Similar research done in other cortical areas, e.g. supplementary and presupplementary motor area (Middleton and Strick, 1997); area 7b of the intraparietal sulcus, and anterior intraparietal area (Clower, West, Lynch and Strick, 2001, 2005), confirms that the outgoing projections from the dentate nucleus are topographically arranged, and reflect functional rather than spatial rela-tionships between the cortical areas that they project to (Strick et al., 2009). Tracing studies done by Borra et al., 2013 show that cortico-pontine projections from the caudal ventrolateral prefrontal cortex involved the dorsomedial and dorsolateral parts of the pontine nuclei (PN). Those parts of the PN also host saccade-related neurons and targets of projections originating from the SC, the FEF, the SEF, and the LIP, and projecting to oculomotor regions of the cerebellar cortex.

2_{The cerebellar module has been defined by Voogd as an anatomical unit consisting of a particular Purkinje cell} microzone with its specific olivary input, together with their innervation of the associated cerebellar or vestibular nucleus.

(30)

Figure 17: In the macaque, transneuronal viral tracers were injected into the arm area of the primary motor cortex (area 4) and prefrontal area 46. Results show the sites of cerebellar cortical terminal label after the injections of retrograde and anterograde tracers (area 4, blue; area 46, green). a) Retrograde projections from area 4. b) Anterograde projections from area 4 to granule cells. c) Retrograde projections from area 46 to Purkinje cells (a, anterior; p, posterior). d) Anterograde projections from area 46 to granule cells. e) Homologous areas in the human cerebellar cortex and a schematic illustration of how they are interconnected with the human cerebral cortex. (Adapted from Middleton and Strick, 2001 by Ramnani, 2006. Reproduced from Ramnani 2006.).

3.2.2 Cortico-olivocerebellar loops

There are two distinct recurrent cortico-rubro-olivocerebellar loops involved with eye movements, the cerebellar components of which are the C2 and D1 Purkinje cell zones, respectively (see Fig-ure 18).

In the first loop, the C2 zone (ansiform lobule, dorsal paraflocculus, and floccular lobe) projects to the posterior interposed nucleus, which is connected with the nucleus of Darkschewitsch in the mesodiencephalic junction, and via the thalamus with the ventral LIP. This region then projects back to the nucleus of Darkschewitsch, which then innervates the rostral medial accessory olive that sends climbing fibres to the C2 Purkinje cells. The other loop runs from the dorsal paraflocculus (D1) to the caudal dentate nucleus, which, like the interposed nucleus, has a mesencephalic and a thalamocortical pathway. The first terminates in the dorsomedial parvocellular red nucleus and the nucleus of Bechterew, and the second in the FEF and LIP. The FEF and SEF project to the pRN, which projects to the principal olive, which again innervates the D1 Purkinje cell zone and sends a collateral projection to the cDN (Voogd et al., 2012).

(31)

Figure 18: Cerebellar connections involved with voluntary saccades and smooth pursuit. Mossy fiber projections from the frontal eye fields are shown in red and from the postrolandic visual areas in green. Cerebello-cortical-olivary climbing fiber loops of the C2 zone and the posterior interposed nucleus are shown in blue, the D1 zone and the caudal dentate nucleus loop in orange; its hatched segments have not yet been verified in primates. Abbreviations: AI anterior interposed nucleus; C2 C2 Purkinje cell zone; D1 D1 Purkinje cell zone; F fastigial nucleus; FEF frontal eye field; MST medial superior temporal area; MT middle temporal area; NRTP nucleus reticularis tegmenti pontis; PEF parietal eye field (lateral intraparietal area); PI posterior interposed nucleus; SEF supplementary eye field; V1, V2 visual areas 1 and 2. (Reproduced from Voogd, Schraa-Tam, van der Geest and De Zeeuw, 2012).

3.2.3 Other cerebellar sub-loops

Besides the larger corticocerebellar loops, there are also several smaller closed subcortical loops involving the cerebellum. Virus transport studies (cf. Bostan, Dum and Strick, 2013) have shown that the dentate nucleus projects disynaptically to the striatum (caudate and putamen). Pro-jections to the striatum originate from motor and non-motor domains in the dentate nucleus. Conversely, the subthalamic nucleus (STN) projects disynaptically to cerebellar cortex, origin-ating from both motor and non-motor domains within the STN, and terminorigin-ating in motor and non-motor regions of the cerebellar cortex. Given this reciprocal connection, the authors hypo-thesize that the putative ‘learning specializations of the basal ganglia (reward-based learning) and the cerebellum (supervised learning) are in some way integrated. Human fMRI studies have shown that activity in the striatum is correlated with reward prediction error in Pavlovian reward asso-ciation tasks (O’Doherty, Dayan, Friston, Critchley and Dolan, 2003). Reward prediction error in these studies is also strongly correlated with cerebellar activity (Peterburs and Desmond, 2016). Within the cortico-olivocerebellar circuits described above there is another loop, running between the inferior olive (IO), the deep cerebellar nuclei (DCN), and the mesodiencephalic junction (MDJ):

(32)

the olivocerebellar mesodiencephalic loop, also known as the myoclonic triangle or the triangle of Guillain-Mollaret. All three elements of this loop are excitatory, and is controlled by local inhibit-ory interneurons in the cerebellar nuclei and MDJ, by the inhibitinhibit-ory feedback from the cerebellar nuclei to the inferior olive, and by the GABAergic Purkinje cell input to the DCN (De Zeeuw et al., 1998; Habas et al., 2010).

Figure 19: Schematic overview of corticocerebellar loops in the oculomotor system. In pink and purple: the two cortico-rubro-olivocerebellar loops, and in blue the olivocerebellar-mesodiencephalic loop.

(33)

4 Discussion: corticocerebellar loops, control theory, and

cognition

Control Theory

In learning a new movement, we first perform it very slowly because it has not yet been prepro-grammed (i.e. learned) in any way. With much cerebrocortical interference and sluggish (40-60ms) proprioceptive feedback, we manage to do it more or less correctly, and subsequently, over many repetitions, achieve unthinking fluency and perfection of the movement. Somewhere along the way, something must have changed in the way the movement is encoded, initiated, and adjusted – the process described above is necessary for conscious learning, but much too slow for rapid, automatic execution (Kawato, Furukawa and Suzuki, 1987; Ramnani, 2006). The model proposed to solve this problem is derived from control theory, a branch of engineering that concerns itself with the behaviour of dynamical systems with inputs, and how that behaviour is modified by feed-back. The basic system consists of a controller (CT) that manipulates a controlled object (CO), an instructor (P) that gives instructions to the controller, and a sensor (SS) that gives feedback to the controller (Ito 2008). In cerebellar control theory, such systems are known as internal models. Internal models are neural representations acquired through learning that can simulate natural processes such as body movements. They encode, update and refine the relationship between a motor command and its consequences each time it is executed. An internal model eliminates the need for external feedback, which is slow and needs to be translated into terms that the motor system can understand (Ramnani, 2006). There are two kinds of internal model that are thought to be at work in the motor system: forward and inverse models. In forward models, an efference copy is made of the motor command signal and is used as an input to the model, which predicts the sensory consequences of an ideally executed movement. This prediction is made parallel to the actual movement, executed by the motor system. The ‘real’ output and the predicted output are then compared to each other: any discrepancy translates as an error signal to the forward model, and is used to update its input-output relationship for better future accuracy (Ramnani, 2006).

In inverse models, the input-output relationship is reversed: the ideal movement is used as input, and the output is a motor command. Rather than model the external world, as the forward model does, the inverse model learns the internal information processing that leads to a motor command. Once learnt, an inverse model can substitute for other brain areas and send a motor command directly to the motor cortex via the VL thalamus. Thus, three systems interact to perfect and streamline a movement as it is learnt: the corticospinal system teaches the forward model until it can replace the external feedback loop; then the forward model teaches the inverse model until it can replace the internal feedback loop, and eventually bypass the process of computing a movement from scratch (Kawato et al., 1987). Due to the rich connectivity between the nonmotor cerebral cortex and the cerebellum, the varied cognitive symptoms resulting from cerebellar lesions, and the uniformity of cerebellar cytoarchitecture, it has been argued with some success that forward

(34)

Figure 20: Forward model. Left, theoretical organization of forward models for motor control. Right, anatomical comparison of the motor control model involving the prefrontal cortex. (Modified from Ramnani, 2006; reproduced from Orozco, 2015).

and inverse models could be used for not only motor, but also cognitive processing (Ito, 2008; Ramnani, 2006).

Neural implementation: corticocerebellar loops

How are these models implemented in the brain? The cortico-ponto-cerebellar projection can serve as a way to send efference copies of motor commands to the cerebellum. Neurons carrying motor commands to the spinal cord collateralize, and the collateral projections synapse onto neurons in the pontine nuclei (Ugolini and Kuypers, 1986). Outputs from the forward model are then relayed back via the thalamus to the area of the cerebral cortex from whence they came; or to the red nucleus, in order to influence spinal mechanisms involved in motor control. The inferior olive is a good candidate for a comparator (Horn, Pong and Gibson, 2004): each dendritic spine of an olivary neuron receives both an inhibitory input from one of the hindbrain regions, such as the cerebellar nuclei and the prepositus hypoglossi, and an excitatory input from the spinal cord, brainstem, mesodiencephalic junction or cerebral cortex. The inferior olive may compare the excitatory as-cending and desas-cending inputs with the inhibitory inputs from the hindbrain (De Zeeuw et al., 1998). In studies done with various types of eye movements, climbing fiber signals were found to represent retinal slips caused by the discrepancy between an executed eye movement and a desired one – i.e., they encoded movement error (Fushiki, Sato, Miura and Kawasaki, 1994). Ito, 2000 identifies long-term depression (LTD) of Purkinje cell activity by climbing fiber signals as the most likely mechanism for motor learning, and contends that over many repetitions of a movement, error signals arising as a result of these movements drive LTD and that the LTD reshapes the cerebellar

Corticocerebellar loops and volitional saccades: evidence for the role of the cerebellum in flexible behavioural control