Removal of information from working memory: A specific updating process

Highlights

• We introduce a method for isolating a process unique to working memory updating.

• Removal of outdated information is an active working memory updating process.

• Removal speed can be measured reliably.

• Removal speed is independent of working memory capacity.

Abstract

Previous research has claimed that working memory (WM) updating is one of three primary central executive processes, and the only one to reliably predict fluid intelligence. However, standard WM updating tasks confound updating requirements with generic WM functions. This article introduces a method for isolating a process unique to WM updating, namely the removal of no-longer relevant information. In a modified version of an established updating paradigm, to-be-updated items were cued before the new memoranda were presented. Overall, longer cue-target intervals—that is, longer time available for removal of outdated information—led to faster updating, suggesting that people can actively remove information from WM. Experiments 1 and 2 demonstrated that well-established effects of item repetition and similarity on updating RTs were diminished with longer cue-target interval, arguably because representational overlap between outdated and new information becomes less influential when outdated information can be removed prior to new encoding. Experiment 3 looked at individual differences, using the reduction of updating RTs to measure removal speed. Removal speed was measured reliably but was uncorrelated to WM capacity. We conclude that (1) removal of outdated information can be experimentally isolated and measured reliably, (2) removal speed is a unique, active WM updating ability, and (3) the view of WM updating as a core executive process that uniquely predicts fluid abilities is overstated.

Introduction

Imagine you ask a colleague for his phone extension and he replies: “It’s 3266. No, hang on, in my new office it’s actually 3257”. Ideally, one should easily discard the last two digits of the outdated information given (i.e., “66”) and replace them in working memory with the correct digits (i.e., “57”). However, this updating of working memory content is no trivial task, and outdated information often continues to affect memory (De Beni and Palladino, 2004, Oberauer, 2001).

Working memory updating has been identified as one of three primary central executive processes Miyake, Friedman, Emerson, Witzki, Howerter, & Wager (2000). Updating has been claimed to be the only executive process to predict fluid intelligence (Chen and Li, 2007, Friedman et al., 2006). However, most updating tasks used in previous research (e.g., Miyake et al., 2000) not only require memory updating but arguably also measure general working memory (WM) abilities. This has led some researchers to conclude that updating tasks constitute reliable assays of general WM capacity (Schmiedek, Hildebrandt, Lovdén, Wilhelm, & Lindenberger, 2009; see also Chuderski et al., 2012, Colom et al., 2008, Martínez et al., 2011).

This creates an unsatisfactory situation. If WM updating tasks measure just the same as other WM tasks such as complex span tasks, then there is no empirical basis for identifying ’updating’ as a separate executive-function factor. Yet, both conceptually and theoretically, updating can be distinguished from maintenance and processing in WM. If updating is to be established as a non-redundant construct, it must be isolated and measured separately from other WM processes.

In a recent individual-differences study, we identified a processing component that was independent of general WM capacity and unique to situations that demanded memory updating (Ecker, Lewandowsky, Oberauer, & Chee, 2010). In that study we analyzed the processing components involved in widely used WM updating tasks, and we identified three separable components: retrieval, transformation, and substitution. The only component process that was unique to WM updating tasks was the substitution of information in memory. To illustrate those components, consider the scenario of a restaurant manager advising a chef early in the evening that they were expecting 20 patrons. If the manager later advised the chef that twice as many guests were expected as before, the chef will need to retrieve the initial expectation (i.e., 20), transform it (i.e., 2 × 20 = 40), and substitute the outdated information with the updated information (i.e., 40). Ecker et al. designed an updating task with eight conditions, fully crossing all possible combinations of retrieval, transformation, and substitution. Applying structural equation modeling to their data, they found that retrieval and transformation operations co-varied with general WM capacity, but that the substitution component did not. This is illustrated by the structural equation model for their updating accuracy data, shown in Fig. 1. This finding was interpreted as showing that substitution is the only process that uniquely represents WM updating, without being “contaminated” by any association with general working memory abilities.

One implication of this analysis is that previous studies measuring WM updating did not separate variance unique to updating from the variance of generic WM processes. As a consequence, the conclusions concerning the predictive relation between WM updating and fluid intelligence (Chen and Li, 2007, Friedman et al., 2006) were arguably not based on a proper measure of WM updating, but may instead reflect the well-known association between higher cognitive functions and general WM capacity (Engle et al., 1999, Oberauer et al., 2005).

In this article, we further decompose the components of WM updating. In Ecker et al. (2010), we suggested that information substitution can be further subdivided into the removal of outdated information and the encoding of new information. For example, the chef would need to remove the number 20 from memory and encode the updated number 40 into the vacant memory slot. Of these two processes, encoding is common to many cognitive tasks. In contrast, we argue that it is the removal process that lies at the heart of, and is specific to, memory updating. Accordingly, we focus on the removal of information from WM. Specifically, we define removal as the unlearning or unbinding of an item from its context. Here we show that (1) removal of outdated information can be separated experimentally from encoding of new information, (2) that removal can be characterized as an active WM updating process, (3) that removal speed can be measured reliably, and (4) that removal speed is independent of WM capacity.

Our conceptualization of removal as an active WM process is inspired by a computational model of working memory, the SOB (“serial-order in a box”) model (Farrell and Lewandowsky, 2002, Lewandowsky and Farrell, 2008, Oberauer et al., 2012). SOB is a two-layer neural network in which items (represented in one layer) are associated to position or context markers (represented in the other layer) through Hebbian learning, which rapidly modifies the matrix of connection weights between the two layers. Memory for items-in-position is hence stored in this weight matrix. Forgetting in SOB is entirely based on interference; there is no time-based trace decay. This assumption is corroborated by a growing body of evidence (Berman et al., 2009, Jalbert et al., 2011, Lewandowsky et al., 2009, Oberauer and Lewandowsky, 2008, Oberauer and Lewandowsky, 2013).

To avoid overloading of the system in the absence of decay, an interference model requires a mechanism to remove outdated information; such a mechanism is implemented in the most recent version of SOB (SOB-CS, a model for the complex span task; see Oberauer et al., 2012). Removal of a specific item involves retrieving that item by cueing with its position marker, and “unlearning” the association between that item and its position. Unlearning is computationally implemented as Hebbian anti-learning. Whereas encoding of information into WM takes less than 500 ms per item (Jolicoeur and Dell’Acqua, 1998, Vogel et al., 2006), removal appears to be a slower process, taking at least about 500–600 ms per item. This estimate is taken from work using a directed-forgetting approach, where participants study two sets of items, one of which is then declared irrelevant before the relevant set is tested. The size of the irrelevant set continues to affect responses to the relevant set for about 1–2 s when the largest irrelevant set comprises three items (Oberauer, 2001; see also LaRocque, Lewis-Peacock, Drysdale, Oberauer, & Postle, 2013). Hence, the time it takes to remove a single item can be estimated as roughly a third of that time, if removal of multiple items occurs sequentially.

Our removal measure is based on the work of Kessler and Meiran (2008), who used a variant of the classic updating paradigm developed by Yntema and Mueser (1962). In this paradigm, individual items (e.g., letters or digits) are presented in a set of individual frames. Items are then repeatedly updated by presenting new items in some frames. On each updating step, at least one item is updated, and participants have to press a key at the end of each step to indicate that they finished updating; this provides an updating RT measure. At the end of the sequence, participants recall the currently memorized set of items.

In SOB, removal of old information and encoding of new information are described as two separate processing steps. It follows that updating should be facilitated if a cue about what information needs to be removed is given ahead of the to-be-encoded new information. In contrast, an incidental feature of standard WM updating tasks to date—including the one used by Kessler and Meiran (2008)—has been that removal cannot commence until the new information is presented for encoding. This is due to the way standard WM updating tasks are designed, namely that the time-point at which people are told what to update coincides with the time-point at which they are given the new information to encode. For example, when the current set of letters in a 3-frame updating task is ‘K-M-R’ and a new letter ‘D’ appears in the third frame, removal of the ‘R’ can only begin when the to-be-updated frame is identified, that is, when the ‘D’ is displayed. Hence updating response times in such a task will include both time for removal and time for encoding. This confound can be avoided by signaling which frame will be updated before the new information is presented. In our new updating task, we thus present cues indicating which items are to be updated before presenting the new to-be-encoded stimuli. People can use the cue only to selectively remove old items from the memory set, not to encode new information. By varying the cue-target interval (CTI) we vary the available time for removal. If people use the CTI to remove the outdated information in the cued frame, then longer CTIs should lead to faster updating RTs.

In Experiments 1 and 2 we show that this is the case, and argue that the reduction of updating RTs by longer CTIs in fact reflects the efficiency of removal during the CTI. Specifically, Experiments 1 and 2 supply an experimental validation of our proposed measure of removal efficiency by (1) showing that there is a big updating RT gain from removing in advance (i.e., a long CTI), and by (2) ruling out alternative explanations for this finding by testing a prediction that is unique to the removal notion, namely that benefits associated with item repetition (Exp. 1) and item similarity (Exp. 2) diminish with a long CTI. This will be followed by Experiment 3, which tests the main hypothesis of our paper, namely that removal efficiency—the one process that is most specific to memory updating—is uncorrelated with WM capacity.

Our first investigation of the effect of the CTI on updating RTs involved testing a specific prediction derived from Ecker et al. (2010). One of the Ecker et al. conditions involved the straightforward substitution of an item; for example, while remembering ‘K-M-R’ in three frames, a ‘D’ might have been presented in the third frame, prompting the participant to update to ‘K-M-D’ (this is the r_no-t_no-S condition in Fig. 1). Another condition was a control condition where an item was presented repeatedly in the same frame on successive updating steps. For example, while a participant was remembering the letters ‘K-M-R’, an ‘M’ might have been presented in the middle frame, prompting a participant to continue to remember ‘K-M-R’ without any need to update the memory representation (this is the r_no-t_no-s_no condition in Fig. 1). We found that repeating an item during the updating task (i.e., maintaining an item unchanged) carried a benefit of roughly 300–400 ms. This benefit should diminish when people are given the opportunity to remove outdated information before encoding the (in this case, identical) updated item, that is, with a long CTI.