A Bayesian inference model for metamemory.

Abstract

The dual-basis theory of metamemory suggests that people evaluate their memory performance based on both processing experience during the memory process and their prior beliefs about overall memory ability. However, few studies have proposed a formal computational model to quantitatively characterize how processing experience and prior beliefs are integrated during metamemory monitoring. Here, we introduce a Bayesian inference model for metamemory (BIM) which provides a theoretical and computational framework for the metamemory monitoring process. BIM assumes that when people evaluate their memory performance, they integrate processing experience and prior beliefs via Bayesian inference. We show that BIM can be fitted to recall or recognition tasks with confidence ratings on either a continuous or discrete scale. Results from data simulation indicate that BIM can successfully recover a majority of generative parameter values, and demonstrate a systematic relationship between parameters in BIM and previous computational models of metacognition such as the stochastic detection and retrieval model (SDRM) and the meta-d′ model. We also show examples of fitting BIM to empirical data sets from several experiments, which suggest that the predictions of BIM are consistent with previous studies on metamemory. In addition, when compared with SDRM, BIM could more parsimoniously account for the data of judgments of learning (JOLs) and memory performance from recall tasks. Finally, we discuss an extension of BIM which accounts for belief updating, and conclude with a discussion of how BIM may benefit metamemory research.

Are the contributions of processing experience and prior beliefs to confidence ratings domain-general or domain-specific?

“…We speculated that metacognitive resolution (e.g., metad'/d'), which has been widely used in previous studies, may not be a suitable measure for investigating domain generality of metacognition, because the relationship between confidence ratings and task performance is affected by specific task requirements (Arbuzova et al, 2021;Samaha & Postle, 2017). In contrast, the parameter P exp in BIM only reflects the contribution of processing experience and prior beliefs to confidence ratings, and is unrelated to task performance (Hu et al, 2021). In fact, this study is just an initial attempt to reduce the confounds introduced by the varied METACOGNITION IN VARIOUS COGNITIVE DOMAINS relationship between confidence and performance across tasks when examining metacognitive domain generality.…”

“…To address this question, Hu et al (2021) proposed BIM, which quantitatively characterize metamemory process based on the dual-basis theory. BIM assumes that the subjective processing experience e for each item is sampled from a distribution of processing experience for each individual, which is correlated with the distribution of objective memory strength m .…”

Optimal Metacognitive Control of Memory Recall

“…Focusing on metamemory in particular, two recent models of judgments of learning are based on signal detection theory (Y. Jang et al, 2012) and Bayesian inference (Hu et al, 2021), both of which have rational bases in probability theory. However, these models only attempt to explain how metamemory judgments are produced, not how they are used-a critical component in a complete rational analysis.…”

“…Like van den Berg and Ma (2018), we jointly consider the problem of how much resource to allocate (here, the resource being time) as well as how to split that resource between items. Like Hu et al (2021), we model metamemory judgments as the product of Bayesian inference about the strength of a memory. And like Suchow and Griffiths (2016), we model the resource allocation problem as an MDP.…”

“…Recently, these phenomena have been understood through computational models of signal detection (Y. Jang, Wallsten, & Huber, 2012) and probability theory (Hu et al, 2021). Less emphasis, however, has been placed on understanding the function of metamemory judgments (Schwartz & Metcalfe, 2017).…”