Toward a Biomimetic Neural Circuit Model of Sensory-Motor Processing

Abstract

Computational models have been a mainstay of research on smooth pursuit eye movements in monkeys. Pursuit is a sensory-motor system that is driven by the visual motion of small targets. It creates a smooth eye movement that accelerates up to target speed and tracks the moving target essentially perfectly. In this review of my laboratory's research, I trace the development of computational models of pursuit eye movements from the early control-theory models to the most recent neural circuit models. I outline a combined experimental and computational plan to move the models to the next level. Finally, I explain why research on nonhuman primates is so critical to the development of the neural circuit models I think we need.

Figure 1:

An image-motion model of pursuit. (A) Target and eye motion during a “step-ramp” target motion, showing the image velocity and accelerations that provide inputs for pursuit. (B) The original three-pathway, image-motion model. The “speed filter” is part of a pathway that is sensitive to image speed. The “onset filter” is part of a pathway that is sensitive only to the onset of motion. The “accel filter” is part of a pathway that is sensitive to smooth image acceleration. (C) Black, red, and green traces show typical eye velocity data and the predictions of the full image-motion model and the model without the acceleration pathway for a step of target speed from 0 to 15 deg/s. (D) Representative time-varying firing rates for an MT neuron during ramp increases and decreases in target speed in an anesthetized monkey. (E) Responses of model MT neurons with the same range of acceleration-deceleration asymmetry found in the data. (F) Decoded target speed and acceleration for a model population of MT neurons with a range of preferred speeds and acceleration-deceleration asymmetries. Figures are modified, with permission from Krauzlis and Lisberger (1989), Lisberger and Movshon (1999), and Lisberger (2015).

Figure 2:

Evidence for control of the gain of visual-motor transmission, role of FEF_SEM, and effect of gain control on pursuit of low-contrast targets. (A) Two-pathway model of pursuit based on the existence of gain control. (B) Speed-tuning curves of an example MT neuron for targets with 100%, 36%, and 12% contrast. (C) Model MT population responses derived from curves in panel G. Open circles show model for high-contrast target motion at 8 deg/s, filled circuits show model for high-contrast target motion at 24 deg/s, and open triangles show model for low-contrast target motion at 8 deg/s. (D) Experimental paradigm to demonstrate gain control. Black arrow points to a brief perturbation of target motion during steady-state pursuit, and red arrow points to the eye velocity response. (E) Black and magenta traces compare responses to the same perturbation presented during steady-state pursuit and fixation of a stationary target. (F) Evidence that FEF_SEM controls the gain of visual-motor transmission. Top trace shows the perturbation presented during fixation. Traces are dashed black trace, response to microstimulation alone in FEF_SEM; dashed red trace, response to perturbation alone during fixation; continuous black trace, response to perturbation during microstimulation; solid red trace, response to perturbation during microstimulation. (G) Speed tuning curves for model MT neurons. Continuous and dashed traces show simulated responses for high- and low-contrast target motion. Figures are modified, with permission, from Egger and Lisberger (2022), Schwartz and Lisberger (1994), and Tanaka and Lisberger (2001). Panel B presents unpublished data from J. Yang and S. G. Lisberger.

Figure 3:

Gain control added to the image-motion model accounts for the effect of reduced dot coherence on the initiation and steady state of pursuit. (A, B) Fast and slow time-base records show the effect of coherence on the initiation of pursuit (A) and steady-state tracking (B). Different shades of green show data for different dot coherences. (C) Image motion model from Figure 1, modified to allow dot coherence to modulate both the gain of visual-motor transmission via G₁ and the gain of eye velocity positive feedback to support steady-state tracking via G₂. (D-F) Effect of dot coherence on the output of the model when modulating both visual and motor gain (D), visual gain only (E), and motor gain only (F). Different colors show model output for different simulated dot coherences. Figure assembled, with permission, from multiple figures in Behling and Lisberger (2020).

Figure 4:

Analysis of trial-by-trial variation in pursuit eye velocity and responses in area MT. (A) Conceptual model of sources of trial-by-trial variation in pursuit and perception. (B) Trial-by-trial correlation of z-scored responses of two MT neurons with a strong noise correlation. (C) Trial-by-trial variation in 17 single-trial examples of the initiation of pursuit. (D) Much smaller trial-by-trial variation shown by the mean and standard deviation of eye velocity during the VOR. (E) Each symbol shows data for a different pair of MT neurons with neuron-neuron noise correlation plotted as a function of the difference between the preferred speeds of the two neurons. (F) Example traces and single trial rasters showing data used to compute MT-pursuit correlations. (G) Each symbol shows MT-pursuit correlation for a single neuron as a function of the target speed relative to preferred speed. Individual neurons appear multiple times for different target speeds. (H) Cartoon graph to illustrate the opposite effects of noise added downstream on behavioral variance and neuron-behavior correlation. Panels are replotted with permission from Lisberger (2010), Huang and Lisberger (2009), and Hohl et al. (2013).

Figure 5:

A biomimetic computational model that accounts for first- and second-order statistics of behavioral and MT neural responses. (A) Schematic of the model; equations available in Egger and Lisberger (2022). (B) Diversity of amplitude and preferred speeds in tuning curves of model MT neurons. (C) Noise correlation among pairs of model MT neurons as a function of the difference in preferred speed of the two neurons. (D) Data showing variance of eye speed at the initiation of pursuit as a function of mean eye speed for targets comprising 2, 6, and 20 deg patches of dots. (E, F) Model predictions for variance of eye speed as a function of mean eye speed for different size targets, with values of 0 (E) or 0.16 (F) for noise in the gain control pathway. (G) Model predictions of eye speed as a function of target speed for three target sizes. (H, I) Model predictions for MT-pursuit correlation as a function of the relationship between target speed and preferred speed with values of 0 (H) or 0.16 (I) for noise in the gain control pathway. Data reproduced with permission from Egger and Lisberger (2022).

Figure 6:

Representation of all components of Bayesian-like behavior in responses of neurons in FEF_SEM. (A) Schematic showing the blend of target speeds in “fast” and “slow” contexts. (B) Evidence of reliability-weighted combination of sensory data and expectations based on previous experience in the initiation of pursuit. (C) Schematic of neural network model used to simulate responses in FEF_SEM. (D) Preparatory activity in data from FEF_SEM neurons and the effect of target speed context on its amplitude, aligned on the end of fixation at t = 1600 ms. (E) Pursuit-related activity in data from FEF_SEM neurons and the effect of target speed context and target contrast, aligned on the onset of target motion. (F) Schematic showing the different measures we used of FEF_SEM activity to obtain the plots in panels I, J, M, and N. (G) Prediction of preparatory and pursuit-related activity by the model circuit and the effect of target speed context and target contrast. (H) Time course of adaptation of amplitude of preparatory activity in switches between fast and slow speed contexts. Black trace shows the prediction of the model; red trace shows normalized data. (I, J) Trial-by-trial correlations of pursuit-related versus preparatory firing rate for data (I) and model neurons (J) for different measures of responses. Each symbol plots data from one real or simulated trial. (K, L) Firing rate modulation during pursuit initiation versus that during preparation for data (K) and model L. (M, N) Distribution across data from neurons of trial-by-trial correlations between pursuit and preparatory firing rate for absolute (M) versus incremental (N) firing rate. All panels reproduced with permission from Darlington et al. (2018).

Figure 7:

A schematic diagram of the anatomical connections within the pursuit system.

Figure 8:

An aspirational model. The model (1) is dynamic, (2) is based largely on computations by model neural circuits rather than decoding equations, and (3) closes the feedback loop so that the sensory responses in MT are determined on the fly.