Neural evidence of the strategic choice between working memory and episodic memory in prospective remembering

Abstract

Theories of prospective memory (PM) posit that it can be subserved either by working memory (WM) or episodic memory (EM). Testing and refining these multiprocess theories of PM requires a way of tracking participants' reliance on WM versus EM. Here we use multi-voxel pattern analysis (MVPA) to derive a trial-by-trial measure of WM use in prospective memory. We manipulated strategy demands by varying the degree of proactive interference (which impairs EM) and the memory load required to perform the secondary task (which impairs WM). For the condition in which participants were pushed to rely more on WM, our MVPA measures showed 1) greater WM use and 2) a trial-by-trial correlation between WM use and PM behavior. Finally, we also showed that MVPA measures of WM use are not redundant with other behavioral measures: in the condition in which participants were pushed more to rely on WM, using neural and behavioral measures together led to better prediction of PM accuracy than either measure on its own.

Keywords: Episodic memory; MVPA; Prospective memory; Working memory; fMRI.

Figure 1

Task diagram and behavioral performance. The dual-task experiment consisted a picture-target detection prospective memory task (“PM”) embedded in an ongoing lexical-decision task (“OG”). Half of the trials were WM_bias trials (1-back lexical decisions and a small set of repeating homogeneous pictures) and half were EM_bias trials (2-back lexical decisions and a large set of trial-unique heterogeneous pictures). Two-thirds of all trials included both tasks (“PM+OG”), and one-third did not require PM responses (“OG only”) Behavioral performance on (A) the PM task in PM+OG trials and (B) the OG task in all trial conditions. (C) Dual-task costs on reaction time in the OG task due to the addition of the PM task (“PM cost”) , and its relationship to PM accuracy across participants. Error bars indicate s.e.m.,*p < .05.

Figure 2

Pattern classification of fMRI data during the delay period predicts PM performance. (A) Voxels that showed significantly greater activity (p < .05, FDR) during probes on PM+OG trials compared to OG-only trials are colored on an inflated atlas brain. This group-level mask was transformed into each participant’s native space and used to mask voxel time series data as input for the pattern classifiers. (B) Trial-averaged classifier evidence for PM trials. PM classifier evidence indicates the difference between target category and distractor category evidence (e.g., “face minus scene” for face-target trials). Error shades indicate +/− 1 s.e.m., interpolated between mean scores from every 2-sec brain scan. Data are not shifted to account for haemodynamic lag. (C) Relating trial-by-trial classifier evidence scores during the delay period (12 to 22 sec) to PM accuracy (hit vs. miss). Data reflect the logistic regression fits (β₁) between PM classifier evidence and PM accuracy. Error bars indicate 95% bootstrap confidence intervals, *p < .05 for 1,000 bootstrap samples.

Figure 3

Classifier evidence scores predict PM accuracy and dual-task PM costs across participants. Higher PM classifier evidence was predictive of (A) better PM accuracy and (B) higher dual-task costs. *p < .05.