Abstract
Humans perceive objects as more similar to recently attended objects than they actually are. This phenomenon, coined serial dependence, is thought to be driven by priors based on the assumption of object permanence. We used fMRI and a linear decoder to investigate how population representations of orientation differed as a function of trial history during a change detection task (Fig1a,b). Behavioral measures indicate that participants showed strong attractive serial dependence (Fig1c, black). Single-trial decoding of fMRI response patterns in visual cortex also revealed strong serial biases (Fig1c, dark-blue) but in the direction opposite to the behavioral bias. The decoded biases were consistent with sensory adaptation whereby populations that were highly active on one trial were less active on the next trial. We hypothesized that readout from early sensory areas accounts for adaptation in normal sensory contexts and that this ‘accounting’ over-compensates in instances of serial dependence (attraction) and under-compensates in instances of behavioral adaptation (repulsion). To simulate this mismatch between expected and true adaptation, we treated observed neural adaptation from passive viewing (approximating ‘natural’ viewing) as the expected adaptation (Fig1c, light-blue) and subtracted this bias from the true neural adaptation observed in our memory task (Fig1c, dark-blue). This residual neural bias was transformed into predicted bias and was found to closely track both the shape and magnitude of the observed behavioral biases (Fig1d, red). The ‘over-adaptation’ model predicts that adaptation on an individual trial should be negatively correlated with serial dependence, a pattern that we directly observed at the level of single trial responses in our data. This model, whereby expected adaptation is accounted for, can be seen as an implicit encoding of the prior belief of object semi-permanence leading to attraction when objects change ‘too-quickly’ (aiding coherent representations) and repulsion when objects change ‘too slowly’ (aiding change detection).