Abstract
It has been reported that damage to V1 can lead to “blindsight”: such patients can discriminate simple visual stimuli above chance and yet allegedly report no conscious visual percept. It has also been suggested that similar forms of “unconscious vision” can be demonstrated in normal observers, i.e. under certain experimental conditions (e.g. low-contrast masked stimuli) subjects can perform above-chance discrimination even when rating confidence/stimulus visibility at the lowest level. However, from a signal detection theoretic (SDT) perspective, these findings may be trivial: subjects may report low confidence/invisibility only because the signal falls below an arbitrary confidence/visibility criterion, rather than being truly invisible. We therefore investigated whether perceptual discrimination without subjective awareness can be demonstrated unequivocally by the rigorous standards of ideal observer analyses. To bypass the criterion problem mentioned above, we probed confidence with a 2-interval forced-choice procedure (Barthelmé & Mamassian, 2009, PLoS CB): normal observers viewed paired intervals of masked stimuli, discriminated the orientation of each (left/right tilted), and then chose which decision they were more confident in (Expt 1) or which stimulus was more visible (Expt 2). Importantly, in one interval the stimulus had zero contrast, meaning above-chance orientation discrimination was impossible. Strong evidence for blindsight would consist of subjects’ discriminating the stimulus above-chance yet failing to consistently select it as more confident/visible relative to the blank interval. Such suboptimal behavior would violate common models of perception. In both experiments, intriguingly, we found that observers judged confidence/visibility with lower sensitivity than they could discriminate orientation, according to SDT measures, suggesting there may be evidence for compromised subjective awareness. Crucially, however, a Bayesian ideal observer model predicts similar behavior. We argue that traditional SDT analyses make unrealistic assumptions about what observers can do in these tasks, and show with simulations how misleading results could be obtained via conventional procedures.
Meeting abstract presented at VSS 2015