The Temporal Architecture of Visual Recognition: OPM-MEG Evidence for Parallel and Hierarchical Processing

Abstract

Visual object recognition is traditionally described as a hierarchical, feedforward process in which information flows sequentially from primary visual cortex through the ventral temporal stream to prefrontal cortex, accumulating over the first 100–150 ms after stimulus onset. An alternative parallel-pathway model proposes that coarse visual information may reach orbitofrontal cortex rapidly via magnocellular-biased projections, allowing prefrontal cortex to generate early categorical predictions before ventral temporal processing is complete. Distinguishing these two architectures requires millisecond-precision measurements of frontal cortex onset latency — a measurement that has been difficult to make reliably with conventional SQUID-MEG, whose cryogenic sensors are positioned approximately 20–30 mm from the scalp and therefore have reduced sensitivity to frontal sources. This thesis presents a single-subject, sensor-level OPM-MEG investigation of visual category processing across six stimulus categories (adult faces, child faces, written words, houses, guitars, and limbs) using a rapid event-related visual category localizer paradigm (725 retained epochs; 0.7% rejection rate). Because OPM sensors mount directly on the scalp, they offer improved sensitivity to frontal signals relative to SQUID-MEG and are well suited to test the temporal predictions of parallel-pathway models. First-significant-onset detection (search window 52–280 ms; ±2 SD threshold) was applied at two levels: single representative sensors (Table 1, Figure 6) and four sensor clusters (occipital n=11; temporal n=14; parietal n=13; frontal n=16; Table 2, Figure 9). Bootstrap 95% confidence intervals (300 resamples) quantify the uncertainty of all onset estimates. The primary finding is that single-sensor ERF onset analysis (Table 1) is consistent with the parallel-pathway model for five of six categories: frontal onset occurs within 13 ms of occipital onset for adult faces, child faces, limbs, guitars, and written words. Cluster-level RMS analysis (Table 2) provides a more conservative picture: frontal onset lags of -8 to +62 ms for five categories are substantially shorter than the 50–100 ms sequential model prediction, with houses showing a principled exception at +174 ms consistent with its dependence on fine-grained spatial analysis. Three independent lines of evidence point toward the same conclusion: ERF onset timing (Figures 6, 9), time-resolved representational similarity analysis with animate-inanimate dissimilarity remaining numerically higher than within-category contrasts in the frontal cluster during the early post-stimulus window (Figure 14), and category-specific frontal beta-band increases for guitars in time-frequency analysis (Figure 13). Together, these findings are consistent with early frontal engagement as predicted by parallel magnocellular pathway accounts, though the single-participant design requires multi-subject replication before any broader generalization.