Abstract
Rapid scene categorization is typically argued to be purely feed-forward. Yet, when navigating in our environment, we usually see predictable sequences of scene categories (e.g., offices followed by hallways, parking lots followed by sidewalks, etc.). Previous work showed that scenes were both easier to recognize, and to discriminate from phase-randomized noise, when shown in ecologically valid, predictable sequences than in randomized sequences (Smith & Loschky, 2019). But, when in scene processing do sequential predictions facilitate scene categorization? We examined this question using EEG. Participants saw scenes in either spatiotemporally coherent sequences (first-person viewpoint of navigating, from, say, an office to a classroom) or their randomized versions. Participants saw 288 scene RSVP sequences (each 1 target and 9 primes), while we recorded their event-related potentials (ERPs). Participants had to categorize one randomly selected target image on each trial, in an 8-AFC task. We found reduced ERP amplitudes for targets in coherent sequences roughly 140 milliseconds after image onset--when ERPs typically first index rapid scene categorization--and during the N300 and N400 components, suggesting both reduced identification costs and semantic integration costs in coherent sequences. Interestingly, such ERP amplitude reductions were predicted by low-level visual similarity between sequential prime-target pairs, suggesting that visual similarity might explain the reduced processing costs in coherent sequences. To test this hypothesis, in Experiment 2, we reran Experiment 1 behaviorally, but replaced the targets with noise images, and asked participants to predict the categories of the missing scenes. Target scenes were more predictable in coherent sequences. Importantly, the correlations of Experiment 2 image predictability with Experiment 1 ERP amplitudes (from 140-450 ms) were greater than for image similarity with ERP amplitudes. Thus, contrary to purely feed-forward accounts of rapid scene category recognition, both predictions for an upcoming scene category and visual similarity between successive scenes facilitate scene “gist”.