Featural representation and internal noise around the visual field

Technical Summary: Featural Representation and Internal Noise Around the Visual Field

Problem Statement
Human visual performance exhibits systematic non-uniformity across the visual field, characterized by two robust polar-angle asymmetries: the horizontal–vertical anisotropy (HVA), where performance is superior along the horizontal meridian compared to the vertical, and the vertical-meridian asymmetry (VMA), where performance is superior in the lower visual field compared to the upper. While these performance fields are well-documented across contrast sensitivity and spatial resolution, the system-level computations that translate non-uniform neural resources (e.g., retinal sampling density, cortical magnification) into these perceptual asymmetries remain largely unexplored. Previous "neural count" hypotheses suggest that performance differences arise solely from variations in the amount of neural tissue representing specific locations; however, image-computable models incorporating only optical and retinal inhomogeneities explain only a fraction of the behavioral variance (approx. 40% for HVA, 10% for VMA). Furthermore, anatomical asymmetries do not specify how neural resources are converted into behavioral decisions. This study addresses the gap in understanding whether these asymmetries arise from differences in featural representation (the perceptual template used to weigh stimulus features) or internal noise (limitations in the decision process), or both.

Methodology
The authors employed a signal-detection task combined with psychophysical reverse correlation and a noisy-observer model to dissociate the contributions of feature weighting and internal noise.

• Experimental Design: Twelve observers performed a peripheral orientation detection task at isoeccentric locations ($6^\circ$) along the horizontal, lower-vertical, and upper-vertical meridians. Stimuli were Gabor patches (horizontal orientation, 2 cycles/degree) embedded in filtered noise (1–4 cpd) or noise-only patches. Signal contrast was titrated to maintain approximately 70% accuracy.
• Double-Pass Consistency: To constrain internal noise estimates, the experiment utilized a double-pass design where identical target stimuli were presented twice within a block (with independently generated distractors). Response consistency (the proportion of identical responses to repeated stimuli) served as a direct behavioral constraint on internal noise.
• Psychophysical Reverse Correlation: The authors estimated observer-specific perceptual templates by analyzing how trial-to-trial fluctuations in stimulus features (orientation and spatial frequency) influenced binary detection responses. This was achieved using a multivariate generalized linear model (GLM) with a probit link, fitted to signal-absent trials to isolate noise-driven decision fluctuations. The resulting weight maps were marginalized to derive one-dimensional tuning functions for orientation gain/bandwidth and spatial-frequency gain/bandwidth.
• Noisy-Observer Model: A computational model was constructed where a decision variable was computed as the inner product of the stimulus energy profile and the estimated perceptual template. This decision variable was corrupted by three potential noise sources: additive noise (private to each pass), multiplicative noise (private and signal-dependent), and shared noise (common across the double-pass pair). The model parameters were jointly estimated by maximizing the likelihood of both trial-wise detection responses and double-pass consistency.

Key Results

1. Behavioral Asymmetries Confirmed: Contrast sensitivity exhibited robust HVA and VMA patterns. However, response consistency, $d'$, accuracy, and reaction times were comparable across locations, suggesting that overall task difficulty or speed-accuracy trade-offs did not drive the asymmetries.
2. Featural Representation (Orientation):
   • At the group level, orientation sensitivity (gain) and selectivity (bandwidth) did not differ significantly between meridians.
   • Crucially, individual differences in orientation sensitivity asymmetry predicted the magnitude of the HVA in contrast sensitivity. Observers with a stronger horizontal-meridian advantage in orientation gain showed a larger HVA.
   • In contrast, orientation sensitivity asymmetry did not predict the magnitude of the VMA.
3. Featural Representation (Spatial Frequency):
   • Spatial-frequency tuning exhibited substantial individual variability, often peaking below the signal frequency (1 cpd vs. 2 cpd signal).
   • Neither spatial-frequency gain nor bandwidth asymmetries reliably predicted the magnitude of either the HVA or the VMA.
4. Internal Noise:
   • Model fitting revealed that a model including additive and shared noise (but excluding multiplicative noise) best accounted for the data.
   • At the group level, additive and shared noise magnitudes were uniform across polar angles.
   • However, individual differences in additive noise asymmetry predicted the magnitude of the HVA. Observers with a horizontal-meridian advantage in lower additive noise showed a larger HVA.
   • Neither additive nor shared noise asymmetries predicted the VMA magnitude.
5. Dissociation of HVA and VMA: The study found that the HVA is linked to systematic variations in orientation sensitivity and additive internal noise, whereas the VMA was not reliably explained by any of the tested featural or noise components.

Key Contributions

• System-Level Characterization: The study moves beyond anatomical "neural count" explanations to characterize the system-level computations (feature weighting and noise structure) that transform sensory input into perceptual decisions.
• Methodological Integration: By combining reverse correlation (to estimate the perceptual template) with a noisy-observer model constrained by double-pass consistency, the authors successfully dissociated the contributions of featural representation and internal noise to visual field asymmetries.
• Dissociation of Asymmetries: The findings provide evidence that HVA and VMA are partially separable phenomena. While HVA is associated with specific computational signatures (orientation sensitivity and additive noise), VMA lacks a clear explanation within the tested framework of feature weighting and decision-stage noise.

Significance and Claims
The authors claim that visual performance fields are not driven by a single visual bottleneck but by location-specific combinations of how the brain encodes relevant feature information and neural noise. Specifically, the HVA reflects systematic variation in the feature weighting of task-relevant orientations and additive internal noise. The study does not claim to have identified the mechanism for VMA, noting that the current task and model may not have been optimized to reveal VMA-linked mechanisms (e.g., those relying on feature selectivity rather than sensitivity).

The framework provides a principled approach to link prevalent perceptual asymmetries to system-level computations. The authors emphasize that their estimated templates and noise parameters are "effective behavioral quantities" that summarize population-level filtering and pooling, rather than direct measurements of single-neuron properties. The results suggest that while non-uniform neural sampling provides a substrate for performance fields, additional readout constraints—specifically in how information is filtered and pooled—determine the final perceptual asymmetry. The study concludes that the HVA is sensitivity-linked, likely due to the horizontal meridian's stronger neural-resource advantages, whereas the computational basis for VMA remains an open question requiring further investigation into mechanisms beyond the scope of the current detection task.