Neural structure of a sensory decoder for motor control

The brain's intricate circuitry transforms sensory information into appropriate motor actions. However, motor outputs exhibit trial-to-trial variability even with identical sensory inputs. This variability follows distinct patterns described by psychophysical laws like Fitts' law (movement variation increases proportionally with speed or size) and the Weber-Fechner law (sensory discrimination thresholds increase with stimulus amplitude). These laws are typically explained by "signal-dependent noise," where noise increases with signal amplitude. Traditional sensory-motor models often employ a "black box" decoding approach, subsuming the transformation process into a single equation with noise incorporated in sensory or motor components. However, this approach struggles to explain variability in complex behaviors and ignores the intrinsic noise within the decoder circuits themselves. This study aims to open the "black box" by investigating the neural computations within the decoder circuits of smooth pursuit eye movements, a well-understood system. Specifically, the research focuses on the role of multiple pathways and pathway-specific noise in shaping the observed variability in motor output.

Existing research on sensory-motor transformations frequently utilizes decoding equations applied to sensory representations to understand brain-generated motor or perceptual behavior. Studies have shown the effectiveness of this approach in various contexts, including visual motion analysis, Bayesian decision making, and sensorimotor decision making. However, the complexity of natural behaviors requires a decoder capable of flexible mapping between sensory inputs and motor outputs. Classic models often attribute variability solely to sensory or motor noise, which becomes insufficient when analyzing complex tasks. The limitations of these models have motivated investigations into the role of noise in the intermediate neural pathways within the decoder.

The study used two male rhesus macaques trained to perform smooth pursuit eye movements while their eye movements were tracked using a scleral search coil. Monkeys were rewarded for pursuing patches of moving dots of varying sizes (2, 6, and 20 degrees) and speeds (4–20 degrees/second). The experiment focused on the initiation phase of pursuit, before eye movement feedback could influence behavior. Data analysis involved identifying and removing saccades, aligning responses by correcting for trial-to-trial variation in pursuit latency, and averaging eye speed within a specific time window. The relationship between behavioral variance and mean eye speed was analyzed for different target sizes and speeds. Signal-dependent noise models were fitted to the data, allowing the Weber fraction to either remain constant or vary with target size. Statistical analyses were performed to determine which model better fit the observed data. A simple computational model was developed to explore the impact of gain noise on the relationship between variance and mean eye speed, and it included three noise sources: sensory noise, motor noise, and gain noise. The quality of fit of models with and without gain noise were compared using the Root Mean Squared Error (RMSE). A more complex, biomimetic model of the pursuit system was then created, incorporating known properties of MT neurons such as direction and speed tuning, receptive field properties (including surround suppression), and noise correlations, and including two parallel pathways for speed estimation and gain control. This model simulated pursuit initiation, and the resulting statistics of model output were compared to actual data from the experiments. Parameters such as surround suppression, threshold nonlinearities, and number of MT neurons were also explored to determine the model's robustness.

The main finding was that target size unexpectedly alters the relationship between behavioral variance and mean in smooth pursuit, violating the predictions of traditional signal-dependent noise models. Specifically, the variance increased more rapidly with mean eye speed for smaller targets compared to larger targets. A signal-dependent noise model allowing for target-size-dependent Weber fractions better predicted the behavioral variance than a model with a constant Weber fraction. Analysis of the data revealed that target size affects the gain of visual-motor transmission, leading to an increase in eye speed with target size for a given target speed. A simple analytical model demonstrated that incorporating noise in the gain control pathway was necessary to explain the observed changes in Weber fractions. A biomimetic model that incorporates parallel pathways for speed estimation and gain control, along with pathway specific noise, reproduced the key behavioral and neurophysiological findings, including the effect of target size on Weber fraction, the magnitude of variance in eye speed, and the trial-by-trial correlations between MT neuron activity and pursuit eye speed. The model's key features were robust across different parameterizations and remained unchanged even when manipulating aspects of the MT neuron model. Results suggest that gain control is subject to independent noise.

The study's findings challenge the traditional view that variability in sensory-motor behavior is solely attributable to sensory or motor noise. The results demonstrate that incorporating noise in the gain-control pathway of a multi-pathway decoder is crucial for accurately capturing the observed behavioral and neural response statistics in smooth pursuit. The two-pathway structure of the model (speed estimation and gain control) is consistent with known neuroanatomy and physiology. This approach provides a more biologically realistic account of sensory-motor behavior than conventional decoding models. The observed target-size effect on Weber fractions was not explained by sensory integration mechanisms. The biomimetic model's success in replicating the complex relationship between behavioral variance and mean, along with neurophysiological data, strongly supports the involvement of pathway-specific noise, particularly in the gain control pathway. The finding has broad implications for understanding noise propagation in complex neural systems and modeling variability in sensory-motor behaviors.

This research provides strong evidence for a novel approach to understanding sensory decoding, emphasizing the importance of considering multiple parallel pathways and pathway-specific noise sources. The proposed model, incorporating noise within the gain control pathway, successfully accounts for a wide range of behavioral and neurophysiological data. Future research could investigate alternative motion pathways, refine the model's representation of MT response properties, and explore the role of other brain regions in shaping pursuit variability.

While the biomimetic model incorporates many known features of the pursuit system, some aspects remain simplifications. For example, the model's representation of MT neuron response properties could be further refined to include more complex surround effects and the potential for feedback to MT. The study focuses on a specific task and species, limiting the generalizability of the findings. Further research is necessary to fully understand how different brain regions contribute to the overall signal and noise of pursuit initiation.