Top-down models explain key aspects of a Speed-of-Sight character recognition task

Abstract

Object recognition is very rapid, typically reaching completion within 150 msec following image onset, consistent with intersaccade intervals in humans. In Speed-of-Sight tasks, recognition can be interrupted by masks presented at a given delay–termed the Stimulus Onset Asynchrony (SOA). Featureless images (black, white, grey or white noise) are minimally effective as masks, even at very short SOAs (e.g. 20 ms). Optimal masks can significantly compromise object identification at SOAs of 60-80 ms or more. We conducted a 2AFC experiment in which subjects reported the location (left/right) of targets presented next to distractors, with both images quickly replaced by identical masks. To limit image parameters while allowing task difficulty to be varied, images were depicted on a 7-segment LED-like display. Targets were always a specific digit (e.g. “2” or “4”). Masks and distractors consisted of digits, letters or non-semantic symbols composed from the same 7 segments. To account for the observed variability in mask efficacy for different target-mask combinations, we constructed a model that combined dynamical variables representing feedforward feature detectors–corresponding to the 7 image segments–with high-level pattern detectors for targets, masks and distractors. Masking was most dependent on feature level competition: the numeral 8 was an effective universal mask whereas the numeral 1 was a poor mask, allowing many targets to be reliably detected after a 20 msec SOA. Accounting for mask effectiveness required postulating top-down or feedback influences from pattern detectors to modulate the confidence or persistence of low-level feature detectors. Our results suggest that masking occurs at the level of low-level features and is strongly modulated by top-down or feedback processes, inconsistent with purely feedforward models often proposed to account for Speed-of-Sight results.

Footnotes

 NSF PetaApps program.