Abstract

Subjective experience at any instant consists of a single (‘unitary’), coherent interpretationof sense data rather than a ‘Bayesian blur’ of alternatives. However, computation of Bayes-optimal actions has no role for unitary perception, instead being required to integrate overeverypossible action-percept pair to maximise expected utility.So what is the role of unitarycoherent percepts, and how are they computed? Recent work provided objective evidence fornon-Bayes-optimal, unitary coherent, perception and action in humans; and further suggestedthat the percept selected is not the maximuma posteriori(MAP) percept but is insteadaffected by utility. The present study uses a binocular fusion task first to reproduce the sameeffect in a new domain, and second, to test multiple hypotheses about exactlyhowutilitymay affect the percept. After accounting for high experimental noise, it finds that bothBayes optimality (MEU) and the previously proposed maximum-utility (MU) hypothesis areoutperformed in fitting the data by a modified maximum-salience(MS) hypothesis, usingunsigned utility magnitudes in place of signed utilities inthe bias function.