Holistic inference explains human perception of stimulus orientation

Perception of stimulus features such as orientation is widely considered a Bayesian inference process. In contrast to previous Bayesian observer models, we propose that perception is a holistic inference process that operates across all levels of the representational hierarchy. We test this hypothesis in the context of a typical psychophysical matching task in which subjects are asked to estimate the perceived orientation of a test stimulus by adjusting a probe stimulus (method-of-adjustment). We present a holistic matching model that assumes that subjects’ responses reflect an optimal match between the test and the probe stimulus, both in terms of their inferred feature (orientation) but also their higher-level (category) representation. Validation against multiple existing psychophysical datasets demonstrates that the model provides a quantitatively detailed and accurate description of subjects’ response behavior including data that previous models fail to even qualitatively account for.