Neural structure of a sensory decoder for motor control

The study addresses how the brain decodes sensory signals to produce motor outputs, focusing on the initiation of smooth pursuit eye movements. Classic psychophysical laws (Fitts and Weber-Fechner) and many models assume signal-dependent noise, typically attributed to sensory or motor stages. However, downstream “decoder” circuits between sensory representations and motor execution also exhibit variability. The key research question is whether noise within specific neural pathways of the sensory-motor decoder contributes uniquely to behavioral variability and how neural circuit structure shapes the variance-mean relationship in behavior. Using pursuit, where area MT provides key sensory drive and links to behavior are well established, the authors test whether changes in target size alter variance-mean relationships in ways incompatible with standard signal-dependent noise and explore a decoder architecture with parallel pathways (speed estimation and gain control) to explain these effects.

Prior work links MT responses to pursuit behavior limits and variability: MT precision matches pursuit precision, MT properties can generate signal-dependent noise, and MT neuron–behavior correlations imply sensory noise influences motor variation and limit downstream contributions. Traditional models often treat the sensorimotor transformation as a black-box decoder with noise placed either in sensory estimates or motor commands. Yet, neural activity in intermediate decoder circuits varies substantially trial-to-trial, suggesting decoder noise may contribute to behavior. Noise correlations in MT can limit information gains with increased population size, challenging simple pooling accounts. Speed perception exhibits approximately constant Weber fractions across target sizes, and motor noise models tie variance strictly to motor command amplitude, predicting no change in Weber fractions for matched mean outputs. FEFsem is implicated in gain control over visual-motor transmission, suggesting a distinct pathway for flexible, context-dependent modulation. These literatures motivate opening the black box to consider pathway-specific computations and noise sources within the decoder.

Behavioral experiment: Two male rhesus macaques performed smooth pursuit of random-dot patches. Patches (2, 6, or 20 deg diameter) moved within an aperture for 150 ms and then translated across the screen for 750 ms. Target speeds were 4, 8, 12, 16, or 20 deg/s in random order; directions were left or right. Eye movements were recorded with scleral search coils with head fixation. Analyses focused on the open-loop initiation interval to avoid visual feedback confounds. Trials with saccades in a 250 ms window from motion onset were excluded. To remove latency variability, an objective alignment procedure inferred pursuit onset and realigned velocity traces. Mean eye speed was computed in 110–190 ms after motion onset for each condition (size, speed, direction), and variance across trials was assessed. Variance-mean relationships were fitted with signal-dependent noise models where Weber fractions were either fixed or allowed to vary with target size; model fits trained on 4, 12, 20 deg/s and tested on 8, 16 deg/s. A bootstrap split-half procedure (1000 iterations) assessed RMSE differences and significance.

Simple analytical model: A tractable decoder model applied a gain G to a sensory estimate s to produce motor output, incorporating independent sensory noise (variance proportional to sensory signal), motor noise (variance proportional to mean motor output), and gain noise (Gaussian with variance σ²). Under independence, the mean output μ = G s and the variance yields an effective Weber fraction Weff that depends on G if and only if gain noise is non-zero. Graphical and analytical exploration tested how gain noise alters variance-mean relationships across gains.

Biomimetic circuit model: A synthetic MT population (N≈1280) with realistic direction and log-speed tuning, receptive field sizes increasing with eccentricity, and center-surround interactions (divisive normalization) was constructed. Preferred directions and speeds were distributed per physiological data; response amplitudes ranged 20–200 spikes/s. Poisson-like noise with empirically constrained pairwise noise correlations depended on differences in preferred direction, preferred speed, and receptive field distance. Single-trial MT population responses were generated for each size-speed condition. A two-pathway decoder processed MT activity: (1) a motion reliability (gain) pathway computed a vector sum to set visual-motor gain G; (2) a speed estimation pathway computed a vector average of log-speed to estimate target speed. The outputs were multiplied to yield simulated eye speed. Gain pathway noise (zero-mean Gaussian, variance σ²) was optionally added. Model parameters were chosen so mean output matched data trends, and the model was evaluated on three criteria: (i) target-size-dependent changes in effective Weber fractions, (ii) overall variance matching behavior, and (iii) realistic magnitudes of MT–pursuit trial-by-trial correlations compared to published data. Robustness analyses varied surround suppression and threshold nonlinearities and assessed effects of population size.

• Changing target size alters the variance-mean relationship of pursuit initiation, breaking standard signal-dependent noise predictions. For a given mean eye speed, variance differed across target sizes; smaller patches (2 deg) showed steeper variance growth than larger patches (6, 20 deg).
• The gain of visual-motor transmission increases with target size: the slope of mean eye speed versus target speed increased with patch size. Slopes (2→6→20 deg) were 0.31→0.52→0.62 in monkey R and 0.46→0.77→0.80 in monkey X; all pairwise increases were highly significant (z-scores > 3.8, p < 0.01).
• Flexible Weber fraction fits (allowing w to vary with target size) predicted held-out data (8, 16 deg/s) better than fixed w fits across monkeys and directions (e.g., Monkey X rightward: RMSE 0.63 vs 0.27 (deg/s)², p = 0.01; Monkey R leftward: 0.35 vs 0.21 (deg/s)², p << 0.01).
• Analytical gain-noise model: With non-zero gain noise (σ² > 0), the effective Weber fraction decreases with increasing gain, reproducing the observed target-size dependence. Without gain noise (σ² = 0), Weff is constant across gains and fails to match data.
• Model comparison on held-out speeds favored gain-noise models: predicted variance RMSE without vs with gain noise—Monkey X left: 0.89 vs 0.56 (deg/s)² (p < 0.01); X right: 0.63 vs 0.44 (p < 0.01); R left: 0.35 vs 0.33 (p << 0.01); R right: 0.46 vs 0.32 (p << 0.01), with bootstrap confirmation.
• Biomimetic two-pathway model required gain-pathway noise to reproduce three empirical statistics: (i) target-size-dependent Weber fraction changes (2 deg showing fastest variance growth), (ii) realistic absolute variance magnitudes in eye speed during initiation, and (iii) realistic MT–pursuit trial-by-trial correlation strengths comparable to reported physiological measurements. Without gain noise, correlations were too large and variance patterns mismatched.
• The two-pathway decoder architecture (vector average for speed estimate; vector sum for reliability-based gain) naturally accounted for mean behavior across speeds and sizes and the observed independence of MT–pursuit correlation signs from preferred speed relative to target speed (contrary to simple vector averaging alone).

The findings demonstrate that behavioral variability in smooth pursuit initiation cannot be explained solely by sensory or motor signal-dependent noise. Target size modulates the gain of visual-motor transmission and, critically, variability arises from noise within the gain-control pathway. Incorporating the known parallel pathways—speed estimation and gain control—into the decoder explains the altered variance-mean relationship and matches both behavioral variance magnitudes and MT–behavior correlations. This reframes decoding from a black-box readout to computation distributed across parallel neural pathways, each contributing distinct noise. The results reconcile prior observations linking MT to pursuit variability while highlighting additional, independent downstream noise sources in gain systems (likely involving FEFsem). More broadly, the study suggests that flexible gain control, central to reliability-weighted estimation and motor optimization, comes with a cost of added variability, which must be accounted for in models of sensorimotor behavior.

The study introduces and validates a decoder architecture for smooth pursuit that mirrors known neural circuitry: parallel pathways for speed estimation and gain control with independent noise sources. A key contribution is identifying gain-pathway noise as necessary to explain how target size changes pursuit Weber fractions and overall variability, while preserving realistic neuron–behavior correlations. By embedding biological structure and pathway-specific noise into decoding, the model provides a parsimonious, mechanistic account of pursuit behavior that surpasses traditional signal-dependent noise frameworks. Future work should directly measure MT responses to these stimuli, further characterize noise correlation structure across features and space, test contributions from other areas (e.g., FEFsem, MST), assess potential feedback gain signals to MT, and explore how gain noise influences broader sensorimotor policies and optimal control models.

Assumptions about MT responses to large patches (surround suppression altering amplitude but not preferred speed) and modeled noise (Poisson-like plus correlated structure) may not fully generalize; extra-classical surrounds in MT are heterogeneous and complex. The decoder’s linear weighting and fixed architecture may deviate from optimal, stimulus-dependent readouts. Observed effects could, in principle, be explained by MT alone if future data reveal stronger noise correlations or spatial integration than currently measured. Other motion pathways (e.g., MST and ocular following circuitry) might contribute to gain changes with larger stimuli. Gain noise could originate from feedback to MT rather than purely feedforward FEFsem pathways. Additional experiments are needed to map pathway-specific noise sources, validate MT responses under these exact conditions, and refine correlation structures.