| Abstract |
Our goal was to construct a biologically motivated neural network model of the primate superior colliculus (SC). SC is a brainstem nucleus involved in the programming and execution of fast eye movements called saccades. Assumptions Our model is based on the following assumptions: 1) It is a neural network model, consisting of units (or elements) that represent groups of neurons, and their activation represent the firing rate of the groups of neurons. 2) It contains among its simulated collicular neurons the layers of motor (TLLBs), visual (Vs), quasivisual (QVs) and predictive visual (PVs) units. 3) It uses the MSH model of the brainstem oculomotor nuclei. This does not mean that it is not compatible with other models of the same structure (e.g. the model by Scudder). 4) All real collicular two-dimensional layers of cells are simulated with one-dimensional arrays of units. This is merely a simplification. The generalization to a two-dimensional model resembling the real SC poses no difficulties. 5) It contains excitatory connections between the output motor elements of the superior colliculus, whose strength varies realistically with the distance between the units of the motor layer. 6) It contains global inhibition at the motor (TLLB) layer. 7) It contains an intracollicular mechanism of local inhibition at the PV layer. 8) It contains a gating mechanism from the substantia nigra to keep the SC silent when the eyes fixate a target. 9) It does not use eye position or eye velocity information as an input to the SC. Instead, a realistic signal originating from the RTLLB neurons of the reticular formation is used as a feedback signal to the PV layer of the SC (and more specifically to inhibitory PVs). The time integral of this signal is directly proportional to the total eye displacement of the eye. 10) It makes no use of a head fixed frame of reference for visual targets either inside the SC or as an input to it. Instead, it uses only the experimentally well justified representations of retinal error and static motor error that are present inside the SC 11) It places the SC outside the feedback loop that is believed to control saccade metrics. 12) It assumes that the electrical stimulation of the deep collicular layers does not directly excite the output of the SC, but instead is stimulates the axons that give input to the motor (output) layer of the SC. 13) The position sensitivity (dependence of the size of an electrically elicited saccade on the initial position of the eye) is a result of the parallel excitation of axons that elicit slow eye movements. We do not deal with the position sensitivity because we assume that it is not related with the collicular mechanism that produces saccades. Results Our model produces the following results: 1) It can produce in cooperation with the MSH model of the burst generator eye movements in agreement with known psychophysics. 2) It produces movements towards the center of gravity of two or more targets (weighted average result). It is thus able to perform a nonlinear vectorial evaluation of its inputs. 3) It always gives one hill of activity at the motor output layer of the SC, with either visual or electrical stimulation. The anatomical position of this activity inside the superior colliculus depends only on movement metrics and not on the number of targets that evoked this movement. The development of this hill of activity is due to the lateral excitatory connections. 4) It contains a retinotopic map of visual targets in the PV and QV layer that is dynamically updated during eye movements (remapping of targets). It thus offers a model for the activation profiles of PVs and QVs. 5) It gives a sequence of two accurate visually guided eye movements (double step saccade paradigm) despite the fact that no visual information is available before the execution of the second saccade. It is shown that the realistic eye displacement signal of reticular origin to the PV layer of the SC is adequate to recalculate saccade metrics in retinotopic coordinates. 6) It gives roughly constant saccade metrics with electrical stimulation, irrespective of current frequency and intensity. 7) It gives a weighted average result when two SC sites are simultaneously electrically stimulated with two electrodes. 8) It gives a sequence of consecutive movements (staircase) with continuous electrical stimulation at a single site. 9) It generates a fixed burst of activity at the output layer of the SC with either a transient visual stimulus or a continuous electrical stimulation. This burst drives the aforementioned MSH model. At prolonged electrical stimulation, it gives a series of bursts at SC exit with latency between them, which give the aforementioned staircase of saccades. 10) It does not simulate the experiment of OPN electrical stimulation. During the residual saccade after release from OPN stimulation, it gives a residual collicular activity, which is the same fixed burst that is elicited for the whole saccade. 11) It reproduces the effects of SC lesions after partial inactivation of its units. It produces smaller than normal (hypometric) horizontal saccades when a region corresponding to bigger horizontal saccades is deactivated, and bigger than normal (hypermetric) horizontal saccades when a region corresponding to smaller horizontal saccades is deactivated. Experimental evaluation of our model All the above assumptions and results are amenable to experimental test. Some experiments have already been done, and are retrodictions for our model or possible falsification for it. Others have not been done yet or remain unresolved as a result of contradictory results from different researchers, and serve as predictions
|
A one dimensional one directional neural network model of the superior colliculus
