Understanding how analysis choices are essential for the meaningful interpretation of visual working memory data

Abstract

Visual working memory (VWM) relies on a distributed cortical network. Yet, the role of individual cortical areas, like early visual cortex (EVC) and intraparietal sulcus (IPS), remains debated. Criteria have been suggested to determine if an area is essential for storage, such as resiliency against visual distraction, and correlations with behavior. Here, we reanalyzed existing data from two independent labs, and caution that adherence to simple criteria could limit our comprehension of the VWM system. Instead, we encourage close consideration of analysis choices for a more complete perspective. When participants remembered an orientation while simultaneously viewing different visual distractors (Rademaker et al., 2019) fMRI activity patterns in EVC and IPS did not distinguish between distractor conditions. While such resiliency implies that both areas are critical for VWM storage, the analysis used (leave-one-out cross-validation) capitalizes on any signal differentiating memory representations during the delay. Instead, an analysis using sensory driven responses for model-training captures only representations that are “sensory-like”, and can yield different conclusions. That analysis choices matter is further illustrated by a task with two memory items – one attended, one unattended (Christophel et al., 2018). Originally, delay-period representations of unattended items were not found in EVC. Our reanalysis reveals that with a model trained on stronger signals (the attended instead of the unattended items) EVC does represent unattended memory items. Finally, both datasets reveal a brain-behavior relationship in EVC, but not IPS. Before declaring EVC essential to storage, a careful examination of analysis choices (like the quantification and read-out of neural error) should be performed to guide interpretation. In sum, a thorough understanding of analyses and the specific principles they test is crucial for unraveling the mechanisms of VWM.