‘Perceptual inference’ refers to the processes that underlie object recognition. Object recognition is no simple feat, since the human brain has no direct knowledge of the outside world, but rather needs to interpret patterns of light that are reflected by objects onto the retina of our eyes. Perceptual inference therefore is akin to deducing the cause of a particular percept. This deductive process is faced with an ‘inverse problem’: there is no unique solution (interpretation) to any one pattern of light that falls onto the retina; in other words, a given pattern of input to the visual system could be caused by a number of different stimulus configurations. This turns the process of perceptual inference into one of weighing up the relative probabilities of a given percept being caused by a number of possible stimuli. How do human observers solve this puzzle? Theoretically, an ideal solution to the problem of perceptual inference is offered by a ‘Bayesian’ statistical framework: in ‘Empirical Bayes’, unknown causes (the so-called ‘posterior probability’) are inferred by integrating the observed evidence (the so-called ‘likelihood’) with a learned and context-sensitive probability distribution of possible causes (the ‘prior probability’). Put simply, knowledge about the context in which a given percept occurs informs its interpretation. A plausible neurobiological implementation of this scheme is provided by Friston’s ‘predictive coding’ model of perceptual inference. Here, each level in the visual processing hierarchy feeds back context-sensitive predictions (priors) about the probable causes of sensation to the next lower level, where they are matched against the incoming sensory data. Mismatches (‘prediction error’) between expected and observed data are fed forward from each level of the hierarchy to the level above. The priors at each level are dynamically adjusted in order to eliminate prediction error at the level below; once all prediction error is eliminated, a stimulus has been ‘recognized’, that is, the system has settled on a unique ‘best guess’ interpretation of the stimulus.

This model, treating the perceptual apparatus as a prediction-generating and -matching machine, offers a stark and intriguing contrast to classical ideas about visual perception, where the visual system was seen as a passive analyzer of bottom-up sensory information. In our Lab, we are interested in testing whether predictive coding mechanisms are indeed the brain’s solution to the problem of perceptual inference (and possibly apply to other decision-making domains as well). We pursue this question by combining functional magnetic resonance imaging (fMRI) with perceptual decision-making tasks, where we can manipulate the sensory evidence (i.e., the ‘likelihood’) as well as the task-context, for instance by varying task instructions or the probability of a particular stimulus category occurring (thus affecting the ‘priors’).