Abstract
For some simple visual tasks, we can make judgments about multiple stimuli in parallel, while for some complex visual tasks we can make only one judgment, showing a complete attentional bottleneck. We asked whether an attentional bottleneck limits performance for a task of intermediate complexity: the semantic categorization of visual objects. Stimuli appeared above and below fixation in rapid serial visual presentation (RSVP), and observers were cued that either one (“single-task”) or both (“dual-task”) objects were relevant. The stimuli were grayscale photographs of isolated nameable objects, and observers judged whether the cued object belonged to a target category (e.g. “animal”). The difference in performance between the single- and dual-task conditions (“dual-task deficit”) was compared to quantitative model predictions. A bottleneck is implemented as an all-or-none serial model (observers can only judge one stimulus at a time, producing a large dual-task deficit). We also considered an independent parallel model (observers can judge two stimuli as accurately as they can judge one) and a fixed-capacity parallel model (a constant amount of information extracted). Results from 9 participants showed a large dual-task deficit (12 ± 1%), most consistent with a bottleneck. Additionally, a bottleneck makes a distinct prediction for responses within a dual-task trial: a given response is more likely to be incorrect when the response about the other stimulus is correct. In contrast, parallel models predict no difference in performance based on the response to the other stimulus. Our data show this difference, consistent with the bottleneck prediction (Δ = 0.04 ± 0.01). We repeated this experiment using a simplified presentation (masking instead of RSVP); results from 6 participants showed the same large dual-task deficit (12 ± 1%). Our findings thus support the presence of a bottleneck for categorization of nameable visual objects, and reject the specific parallel models considered here.