Abstract
The N-back task is one of the most popular paradigms for studying the cognitive mechanisms of working memory (WM). The task requires the observer to view a sequence of stimuli and judge whether the current stimulus (probe) matches the one presented N stimuli ago (target). A key phenomenon is that the intervening stimuli (distractors) interfere with task performance. Unfortunately, the classic N-back task uses complex categorical stimuli, making it unfit to quantify the effect of feature similarity on interference strength. Here, we introduce the "analog N-back task", which utilizes stimuli varying continuously in orientation or color. This task variant enables us to measure interference strength on a continuum, providing data suitable for identifying the sources of interference using computational models. In the analog 2-back task, we found that interference increased with feature similarity between the probe and both task-relevant (1-back) and task-irrelevant (3-back) distractors. We next developed and evaluated three main models that each incorporated a Bayesian decision step and differed from an optimal non-interference model in one component only: an early-pooling model, a late-pooling model, and a substitution model. Model comparison suggests that interference emerges late in processing, most likely due to confusion between stimuli during WM retrieval. Our work puts the study of interference in the N-back task on a firmer computational footing and provides a unified framework for examining the sources of interference across domains.