Basal Ganglia: Control of Saccades
The basal ganglia are a set of subcortical interconnected nuclei, subdivided in parallel circuits. They form loops with specific cortical and subcortical regions and play a central role in the selection of actions as well as in learning to bias these selections towards the most profitable options. One of these loops appears to be specialized in saccadic eye movements. Most of the existing computational models consider that it is responsible for the choice, among all possible targets, of the target of the next saccade.
Basal Ganglia Saccadic Circuitry
The basal ganglia are a set of subcortical nuclei common to all vertebrates (see “Basal Ganglia: Overview”). The basal ganglia are components of two main types of loops: cortico-baso-thalamo-cortical loops (Alexander et al. 1986) as well as subcortical ones (McHaffie et al. 2005). Within these two categories, sub-loops can be characterized, dedicated to various functions; one cortical loop dedicated to eye movements has been identified, as well as a few subcortical ones, all projecting to the superior colliculus, where information converges to allow the generation of eye movements (see “Oculomotor Control, Models of”). The cortical loop receives contributions from the frontal eye fields, the supplementary eye fields, the dorsolateral prefrontal cortex, and the parietal cortex and goes through the ventral anterior, ventrolateral, and medial dorsal thalamic nuclei. The subcortical ones all receive inputs from the superior colliculus via the lateral posterior, the intralaminar, and the pulvinar nuclei of the thalamus.
The outputs of the basal ganglia are inhibitory and tonically active, so that at rest, they maintain their targets under constant inhibition, preventing their activation. The generic role of the basal ganglia is understood as an action selection system, used to solve resource allocation problems (Mink 1996; Redgrave et al. 1999): competing channels within the basal ganglia represent the different potential actions, and when simultaneous inputs activate different channels, the basal ganglia resolve the competition and disinhibit only one of the outputs and thus allow the activation of a single action. In the context of saccadic eye movements, at a given moment, the activities caused by multiple points of interest present in the visual field are susceptible of eliciting multiple saccades, which would allow for the refined analysis by the fovea of these points of interest. The role of the basal ganglia is thus understood as filtering these multiple solicitations in order to keep only the one with the highest priority.
The basal ganglia also interact with dopaminergic nuclei (the ventral tegmental area and the substantia nigra pars compacta): the basal ganglia project to these nuclei and receive modulatory projections at the striatal, pallidal, and subthalamic levels. The current theory proposes that these interactions are the neural substrate of reinforcement learning. They would allow learning, by trial and error, of which actions are the most profitable and would bias the selection process accordingly. The dopaminergic signal is thought to correspond to a measurement of the reward prediction error, which is used to modify the strength of the cortico-striatal synapses according to temporal-difference learning algorithms (see “Reward-Based Learning, Model-Based and Model-Free”). In the saccadic eye movement context, this would be the substrate of the evaluation of the priority of the competing targets, allowing the brain to make the best decision according to experience.
Electrophysiological recording in monkeys showed that saccade-related activity in the basal ganglia of monkey is quite rich (Hikosaka et al. 2000). Input neurons in the striatum exhibit visual and motor saccade-related activity, which is also reflected in the output activity of the substantia nigra. But more interestingly, one third of the neurons have a working memory-specific activity: some of them have visual or motor burst activity only for memory-driven saccades, while the rest have sustained activity while the position of a target is kept in working memory.
Computational Models of the Basal Ganglia Saccadic Circuitry
Relatively few computational models of the basal ganglia have specifically targeted the saccadic circuitry (Dominey and Arbib 1992; Dominey et al. 1995; Brown et al. 2004; Chambers et al. 2005; N’Guyen et al. 2010, 2014; Guthrie et al. 2013; Thurat et al. 2015; Cope et al. 2017; James et al. 2018); thus no final consensus about the role and the specific mechanisms of the basal ganglia saccadic circuit has been reached. All of them have been built at the same level of description, using variations of rate-coding models of neuron populations.
A specificity of models of the saccadic basal ganglia, compared to other basal ganglia models (see “Basal Ganglia: Overview”), is that the targeted motor systems (namely, superior colliculus and brainstem saccade generators) are relatively well known and have already been extensively modeled. Consequently, most of the aforementioned models include motor components allowing to simulate actual behavior.
All of these models consider that the neurons that are involved in saccades towards visible targets belong to a circuit in charge of selecting the target of the next saccadic movement in a winner-takes-all manner, in line with the current understanding of the general basal ganglia function (see “Basal Ganglia: Mechanisms for Action Selection”). Those models are based on traditional basal ganglia computational models, adapted to saccade selection: each channel in competition codes for a given saccade metric, and the selected channel disinhibits the area of the superior colliculus specifically encoding that metric. The common selection mechanism among those models is based on the focused inhibition of the striatum, by which each channel tries to inhibit its output, and the diffuse excitation of the subthalamic nucleus, used to excite the competitors.
The models which also incorporate reinforcement learning (Dominey et al. 1995; Brown et al. 2004; N’Guyen et al. 2010, 2014; Guthrie et al. 2013) rely on variations of classical temporal-difference reinforcement learning algorithms (see “Reinforcement Learning in Cortical Networks”). In these models, derived from machine learning theories, the dopamine released in the striatum is supposed to correspond to the reward prediction error signal used by these algorithms; it drives the modifications of the strength of the cortico-striatal synapses.
Notably, in the Dominey and Arbib (1992) model, a second loop parallel to the “visual” one is dedicated to working memory. It is supposed to select which of the current potential targets have to be stored, in order to allow for saccades towards targets that may disappear from the visual field. Such a functionality requires a remapping capability: as the frontal eye field, parietal cortex, or superior colliculus maps, which are supposed to store these memories, are retinotopic, the positions of the next saccades have to be updated after each eye movement. This mechanism can be extended to store the order of the targets, allowing for a more elaborated sequence generation capability (Dominey et al. 1995). Note that models of the role of the basal ganglia in working memory have been proposed in other contexts and could easily be adapted to the saccadic function (see “Working Memory, Models of”).
Concerning the subcortical basal ganglia loops through the superior colliculus, no model has ever considered the cohabitation and the possible interactions of the three anatomically described circuits (McHaffie et al. 2005). Those which considered the basal ganglia-superior colliculus loops included only one of them. Earlier models used it in a slave mode (Brown et al. 2004), where it executes the motor decisions taken by a master cortical loop; more recently it has been considered as an additional player operating in parallel to the cortical loops (Chambers et al. 2005; N’Guyen et al. 2010, 2014; Cope et al. 2017; James et al. 2018), and its possible ability to autonomously make decisions hasn’t been much explored yet (Thurat et al. 2015).
- Chambers JM, Gurney K, Humphries M, Prescott TJ (2005) Mechanisms of choice in the primate brain: a quick look at positive feedback. In: Bryson JJ, Prescott TJ, Seth AK (eds) Modelling natural action selection: proceedings of an international workshop. AISB, Sussex, pp 45–52Google Scholar
- Cope AJ, Chamber JM, Prescott TJ, Gurney KN (2017) Basal ganglia control of reflexive saccades: a computational model integrating physiology anatomy and behaviour. bioRxiv preprint. https://doi.org/10.1101/135251
- N’Guyen S, Thurat C, Girard B (2014) Saccade learning with concurrent cortical and subcortical basal ganglia loops. Front Comput Neurosci 8(00048). https://doi.org/10.3389/fncom.2014.00048