Abstract
Information in visual working memory is influenced by irrelevant past information such that reports show a bias towards the remembered feature values of previous trials (Fischer & Whitney, 2014). Here, we propose that such serial dependence can be understood by considering working memory representations as distributed (population-based or probabilistic representations; e.g., Schurgin et al. 2018; Bays, 2014), rather than as point representations. In this framework, the previous trial causes lingering activity across all sensory channels (e.g., all colors) that influence the current trial. This predicts how both the strength of the activity from the previous and current trial as well as the population overlap (i.e. feature similarity) will affect the degree of serial dependence.
We used experiments with both color and orientation to test these predictions. Fifty participants were asked to reproduce a single stimulus after a short delay. On each trial, we induced either a high or low strength memory via encoding time/masking manipulations. The order of trials was such that memory strengths and the distance in feature space between subsequent stimuli was balanced within-subject.
As expected, memory strength of the current trial affected dependence on the previous trial such that this dependence was larger for low memory strength trials. We also found an interaction between the current trial strength and the feature similarity between subsequent trials: dependence on the previous trial for strong memory trials only occured when the subsequent stimuli were dissimilar; conversely low memory strength trials showed stronger dependence on the previous trial when subsequent stimuli had high feature similarity.
These experiments suggest that serial dependence can be explained by a mechanistic account of visual working memory that represents both memory strength and feature similarity in terms of distributed population activity.