Abstract
Current dominant theories of perception hold that when faced with multiple sources of information, the human brain weights them by their respective reliability in a near-optimal fashion. Here we present results to show that such principles, though attractive from an engineering perspective, may break down in humans in conditions as common as attentional cuing. In order to account for the finding that attention leads to relatively conservative perceptual decisions, we recently proposed a signal detection theoretic model (Rahnev et al 2011 Nature Neuroscience) which assumes a single read-out mechanism compromising between the different variance structures associated with attended and unattended stimuli. Specifically, the observer sets a common decision criterion for both the attended and unattended, and thus is systematically suboptimal for both. One counter-intuitive prediction of the model is that when subjects try to incorporate prior expectations (e.g. the upcoming grating is 70% likely to be tilted to the left rather than right) into their perceptual decisions, a relatively strong weight would be given to the prior expectation when the stimulus is attended and reliably processed, rather than unattended and poorly processed. This prediction goes against principles of Bayesian optimality, and thus allows for a strong test for our model. We confirmed our a priori predictions. Results from Experiment 1 suggest that perhaps a strong weight is given to the prior when the stimulus is attended, because attention leads to relatively conservative subjective decision confidence compared to inattention, which may bias the weight given to the prior. However, Experiment 2 shows that even when the attended stimulus led to both a stronger objective informational strength (higher d’) as well higher decision confidence compared to inattention, subjects still assigned higher weight to the prior. These results pose strong constraints on models of the read-out mechanism in perceptual decisions.
Meeting abstract presented at VSS 2012