Violations of Conditional Independence in Mixed-State Models of Visual Working Memory

Abstract

Mixed-state models of visual working memory assume discrete mental states ofknowing and guessing; information is either represented in memory or not at all. At the core of most conceptualizations of mixed-state models is the notion of conditional independence, that is, the probability of a process taking place when a given memory state is reached is independent of the probability of reaching said state. For example, the probability of committing a memory-based error being conditionally independent of the probability of a memory-based response being made. We test this assumption using an extended recognition-memory paradigm that allows for second choices, along with a task-difficulty manipulation. The second choices obtained through this paradigm provide testable predictions for conditional independence that are violated in the reported data. Overall, our results indicate that the probabilities of memory-based responses and of memory-based errors go hand in hand.

Zooming in on what counts as core and auxiliary: A case study on recognition models of visual working-memory

“…This finding is consistent with results obtained in visual delayed estimation paradigms, which no longer consider the classic all-or-none discrete-slot model of memory as a plausible model of visual working memory (van den Berg et al, 2014;Zhang and Luck, 2008). Support for continuous resource models of visual working-memory also corroborates results from a recent minimal assumptions critical test (Winiger et al, 2021), recent ROC analyses using confidence-based judgements (Robinson et al, 2020;Williams et al, 2022), and novel unifying framework of memory (Schurgin et al, 2020). Independent of the evidence we find favoring the continuous resource model, at a minimum, our reanalysis shows that strong claims of support for discrete-slot models from the Rouder et al (2008) and Donkin et al (2014) papers are not justified.…”

“…Researchers should also use a conceptual analysis to identify the full scope of auxiliary assumptions that can be made at multiple levels of a theory (Figure 2A). These assumptions may be theory-specific, meaning that they are used to parameterize specific instantiations of a given theory (Kellen et al, 2021). Likewise, these assumptions may be theory-general and bare on predictions regarding how variables change as a function of a manipulation that is not central to the hypothesis at hand, but requisite for testing the hypothesis in practice.…”