Abstract
A crucial feature of visual working memory is its flexibility: we are able to hold any type of visual stimulus in mind. This has prompted the development of biologically-realistic network models of working memory that account for both flexibility and capacity limits. Of particular interest to us is a model by Bouchacourt and Buschman (2019). This model has a two-layer structure, with one layer of specialized, sensory neurons (representing early visual cortex) that project to a flexible, “random” layer (representing higher-level regions). Critically, interference in this model is instantiated in the random layer but not directly in the sensory layers. A resulting behavioral prediction is that holding a heterogeneous set of items in mind (e.g. colorful circles and oriented lines) should produce comparable interference to a homogeneous set of equivalent size (e.g. colorful circles only) because competitive interactions occur in less specialized networks. We tested this prediction using a visual working memory task where participants (N = 21) held two or four items in mind. These trials were either homogeneous (all colors or all orientations) or heterogeneous (half colors and half orientations). One item in the array was probed, and participants reported the item’s color or orientation using the corresponding color or orientation wheel. For each participant and experimental condition, we calculated the circular mean and standard deviation of the response error. For both color and orientation responses, we found that 4 item trials produced more interference than 2 item trials, and, critically, that homogeneous trials produced more interference than heterogeneous trials. These data suggest that competition in a random network layer may be insufficient to explain interference between memory items and that, instead, these competitive interactions may occur jointly in early and later stages of processing.