Dynamics of efficient ensemble coding

Ensemble coding creates compressed representations of a stimulus array. However, not all items in the array necessarily contribute equally to the ensemble code. For example, when discriminating the ensemble average against a reference, items whose feature values lie closer to the reference are typically weighted more strongly. We have recently shown that this inhomogeneous weighting can be explained as a form of efficient coding, by which the precision of the sensory representation is dynamically adapted according to the array's overall statistics relative to a variable reference. However, the specific process underlying the formation of such a relative efficient ensemble code remains unknown. Here, we probed the dynamic interplay between the presentation of the ensemble and the reference stimuli. We found that the relative timing between the presentation of the reference and the stimulus ensemble has a strong effect on participants' decision behavior. Model analysis showed that efficient ensemble coding is established only when reference and ensemble are simultaneously presented. It is much weaker when the ensemble preceded the reference, and is largely absent when the ensemble followed the reference. As captured by our model, reduced efficient ensemble coding coincides with decreased decision accuracy in those asynchronous conditions. Our results indicate that any temporal offset between the ensemble and reference stimuli substantially disrupts the dynamic reallocation of coding resource. This suggests that efficient ensemble coding is not the result of a preparatory attentional process nor is it due to evidence selection at the decision stage. Rather, it arises from a fast interaction between the simultaneously evoked, sensory representations of reference and ensemble stimuli.