Is working memory domain-general or domain-specific?

Abstract

Given the fundamental role of working memory (WM) in all domains of cognition, a central question has been whether WM is domain-general. However, the term 'domain-general' has been used in different, and sometimes misleading, ways. By reviewing recent evidence and biologically plausible models of WM, we show that the level of domain-generality varies substantially between three facets of WM: in terms of computations, WM is largely domain-general. In terms of neural correlates, it contains both domain-general and domain-specific elements. Finally, in terms of application, it is mostly domain-specific. This variance encourages a shift of focus towards uncovering domain-general computational principles and away from domain-general approaches to the analysis of individual differences and WM training, favoring newer perspectives, such as training-as-skill-learning.

Keywords: brain training; domain-generality; neural correlates; resource models; working memory.

Figure 1.. Theoretical and empirical aspects of resource division in WM.

(a-d) Schematics of four accounts of resource division when one item (left) or four items (right) are to be stored, adapted from [12]. (a) Slot model (k = 3). (b) Slots + averaging (SA) model. (c) Continuous resource (equal precision) model. (d) Continuous resource (variable precision) model. Distributions under each figure show error around the true feature value of the probed item. This distribution is similar across models when one item is to be stored (hence shown once). When the number of items increases beyond capacity, all models predict increased error. In slot models, this distribution is a mixture of high-precision responses (white) and random guesses (yellow); in SA model, a mixture of low-resolution recall (white) and random guesses (yellow). In equal-precision continuous resource model, error increases continuously with the number of items, and in variable-precision, the distribution is an infinite mixture of distributions with varying error rates. (a) and (b) are mixed-state models. (e) Ratings of tokens in the perceptual task. (f) Changes to mean deviation scores in the perceptual task as a function of number of items and position. (g) The cueing paradigm. Uncued = non-cued items on a cued trial. Baseline (not shown) had no cues. (h) Mean deviations scores in (g). (i) Schematic of the activation in the language production system for target “cat” (Box 1). Orange indicates activation through feedback. (j) Example trial and ALINE coding in the production task. (k) Mean ALINE distance as a function of word length in (j).

Figure 2.. Neural correlates of phonological and semantic WM.

(a) Paradigm for testing phonological and semantic WM. (b-d) ROI-based results of representational similarity analysis in (b) Left superior temporal gyrus (STG), (c) left supramarginal gyrus (SMG), and (d) left angular gyrus (AG). Error bars represent the standard error of the mean. Dashed lines indicate the typical boundaries between the encoding period and the delay period. pho: phonological; sem: semantic. Asterisks indicate the significance of one-sample t-test: *p < .05, **p < .01, ***p <.001. Adapted from [64]. (e) Beta values of the regions significantly associated with decreased performance in phonological and (f) semantic WM after accounting for lesion volume, input processing, and the respective opposing WM task (p values < 0.05); adapted from [71].

Figure I.. Schematic of potentially domain-general and domain-specific regions in WM.

GPe = globus pallidus, I = input, O = output, PFC = prefrontal cortex, SNr = substantia nigra. Adapted from [113].

Figure I.. Flexible WM model [103].

(A) Model layout. The sensory network is composed of 8 ring-like sub-networks. The inset shows center-surround connectivity within a sensory sub-network. The connections to the random network are randomly assigned and balanced. (B) Raster plot of an example trial with 8 sensory sub-networks (512 neurons each) randomly connected to the same random network (1,024 neurons). Six sensory sub-networks (1–6) receive a Gaussian input for 0.1 s during the “stimulus presentation” period (shaded blue region). Representations are maintained (without external drive) for four of the inputs. Reproduced from [103].