The Development of Real-Time Stability Supports Visual Working Memory Performance

The Development of Real-Time Stability Supports Visual Working Memory Performance: Young Children’s Feature Binding can be Improved

Through Perceptual Structure

Vanessa R. Simmering University of Wisconsin–Madison

Chelsey M. Wood University of Iowa

Working memory is a basic cognitive process that predicts higher-level skills. A central question in theories of working memory development is the generality of the mechanisms proposed to explain improvements in performance. Prior theories have been closely tied to particular tasks and/or age groups, limiting their generalizability. The cognitive dynamics theory of visual working memory development has been proposed to overcome this limitation. From this perspective, developmental improvements arise through the coordination of cognitive processes to meet demands of different behavioral tasks. This notion is described as real-time stability, and can be probed through experiments that assess how changing task demands impact children’s performance. The current studies test this account by probing visual working memory for colors and shapes in a change detection task that compares detection of changes to new features versus swaps in color-shape binding. In Experiment 1, 3- to 4-year-old children showed impairments specific to binding swaps, as predicted by decreased real-time stability early in development; 5- to 6-year-old children showed a slight advantage on binding swaps, but 7- to 8-year-old children and adults showed no difference across trial types. Experiment 2 tested the proposed explanation of young children’s binding impairment through added perceptual structure, which supported the stability and precision of feature localization in memory—a process key to detecting binding swaps. This additional structure improved young children’s binding swap detection, but not new-feature detection or adults’ performance. These results provide further evidence for the cognitive dynamics and real-time stability explanation of visual working memory development.

Keywords: working memory development, visual feature binding, computational model

Working memory is a fundamental cognitive process that un- derscores performance on a range of tasks, from following a teacher’s multistep instructions to planning a route through a grocery store. This ability to hold past information in mind to use flexibly in service of behavior correlates with scholastic achieve- ment (e.g., Cowan et al., 2005; Pickering & Gathercole, 2004; Raghubar, Barnes, & Hecht, 2010). Longitudinal studies show that infants’ visual memory predicts their higher-level cognitive skills up to 10 years later (e.g., Rose, Feldman, & Jankowski, 2012), suggesting that the foundations of adaptive cognitive functioning may emerge very early in development. However, the processes

that are shared among different measures of memory across tasks and age groups remain unknown (Simmering, 2016).

Working memory development has been addressed somewhat independently across domains, with different explanations of im- provement on simple tasks (i.e., requiring only maintenance) or complex tasks (i.e., requiring manipulation or task switching). In complex tasks like reading or counting span, which require pro- cessing new information while maintaining prior information, de- velopmental improvements likely arise through increases in processing speed, rehearsal, and/or resistance to interference. However, simple tasks that rely less on such processes still show

This article was published Online First June 19, 2017. Vanessa R. Simmering, McPherson Eye Research Institute, Waisman

Center, and Department of Psychology, University of Wisconsin–Madison; Chelsey M. Wood, Department of Psychology, University of Iowa.

Thanks to the families who participated in this research and the research assistants who aided in data collection. The work at the University of Iowa was made possible by the National Science Foundation (HSD 0527698). A subset of the data in Experiment 1a were part of the second author’s senior honors thesis, and was presented at the 68th Biennial Meeting of the Society for Research in Child Development, Denver, CO, and the 31st Annual Meeting of the Cognitive Science Society, Amsterdam, Nether- lands. Data from Experiment 2 were presented at the 45th Annual Meeting of the Jean Piaget Society in Toronto, Ontario. Participant recruitment and

programming at the University of Wisconsin–Madison was funded by the Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health (R03-HD067481) and the Waisman Intellectual and Developmental Disabilities Research Center grant (P30HD03352). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the National Science Foundation. Thanks to Rob Olson for programming assistance, and to Jeffrey S. Johnson and John P. Spencer for input during the conceptualization of the project.

Correspondence concerning this article should be addressed to Vanessa R. Simmering, Department of Psychology, University of Wisconsin–Madison, 1202 West Johnson St., Madison, WI 53706. E-mail: simmering@wisc.edu

T hi

s do

cu m

en t

is co

py ri

gh te

d by

th e

A m

er ic

an Ps

yc ho

lo gi

ca l

A ss

oc ia

tio n

or on

e of

its al

lie d

pu bl

is he

rs .

T hi

s ar

tic le

is in

te nd

ed so

le ly

fo r

th e

pe rs

on al

us e

of th

e in

di vi

du al

us er

an d

is no

t to

be di

ss em

in at

ed br

oa dl

y.

Developmental Psychology © 2017 American Psychological Association 2017, Vol. 53, No. 8, 1474–1493 0012-1649/17/$12.00 http://dx.doi.org/10.1037/dev0000358

1474

 

 

parallel developmental improvements, and relate reliably to higher cognition (e.g., Cowan, 2013). Thus, it is critical to understand performance and development even in simple memory tasks.

As a step in this direction, Simmering (2016) proposed the cognitive dynamics theory of visual working memory (VWM) development. Although this theory was developed to address in- creases in VWM capacity, the mechanism was proposed to be a general account of development that can be applied across do- mains (see General Discussion section). In particular, Simmering (2016) posited that increases in real-time stability could explain developmental improvements across tasks. In contrast to long-term notions of stability, in which earlier behavior predicts outcomes later in development, real-time stability characterizes how effec- tively the memory system functions in response to different de- mands of behavioral tasks. Through implementation in a compu- tational model, this developmental mechanisms has quantitatively captured performance on VWM tasks from infancy (5 to 13 months; Perone, Simmering, & Spencer, 2011; Perone & Spencer, 2013b, 2014), early childhood (3 to 7 years; Simmering, 2016; Simmering, Miller, & Bohache, 2015), and adulthood (e.g., John- son, Simmering, & Buss, 2014).

Here we test the real-time stability hypothesis further, first by showing that modifying the behavioral task can reveal instability early in development, and second by augmenting stability in the moment of the task. Specifically, we tested children’s and adults’ memory for multifeature visual stimuli (color-shape conjunctions), with particular attention to how memory is probed. These studies are not intended to contrast theories of multifeature representation in VWM, but rather leverage this paradigm to test a proposed developmental mechanism. In the sections that follow, we describe theories of VWM development from infancy and middle child- hood. We then briefly review accounts of feature binding in VWM, which have focused primarily on adults. Lastly, we com- bine theoretical frameworks from these domains to generate and test two specific predictions.

Theories of Visual Working Memory Development

Research on VWM development can be divided into two bodies of work: one addressing infancy (see Reznick, 2009, for review) and one focusing on middle childhood and adults (see Cowan, 2016, for review). We briefly review the dominant findings and theories in these two areas in turn. Infant studies use looking and reaching paradigms due to infants’ limited behavioral abilities. Across these methods a general pattern has emerged (see Reznick, 2009; Rose, Feldman, & Jankowski, 2004, for reviews): over development, infants form memory representations more quickly, maintain representations across longer delays, use representations more robustly in recognition or search, and represent more com- plex information. Theories explaining these changes focus on development of neural circuits supporting VWM (e.g., prefrontal cortex, Diamond, 1990; Ross-Sheehy, Oakes, & Luck, 2003; or parietal regions, Oakes, Messenger, Ross-Sheehy, & Luck, 2009) and related cognitive consequences (e.g., inhibition, Diamond, 1990; or object individuation, Oakes et al., 2009).

Studies of VWM during early to middle childhood typically use tasks adapted from the adult literature, such as the change detec- tion task (e.g., Luck & Vogel, 1997). In this paradigm, a memory array is presented with a small number of simple objects, then

following a short delay, a test array is presented in which the items either match the memory array or one has changed. Participants indicate whether the memory and test arrays were “same” or “different.” Most developmental studies have tested memory for colors (e.g., Cowan et al., 2005; Isbell, Fukuda, Neville, & Vogel, 2015) or shapes (e.g., Simmering et al., 2015), and have found gradual increases in performance with age (see Simmering & Perone, 2013, for review). These results are commonly interpreted as evidence for a developmental increase capacity—the number of items that can be held in memory at once.

This interpretation aligns with a “slot”-like characterization of VWM, in which capacity is conceptualized as a fixed number of discrete representations that can be encoded (see Suchow, Foug- nie, Brady, & Alvarez, 2014, for review). Most theories of VWM development explicitly or implicitly endorse this perspective. For example, Cowan has proposed that the number of “chunks” that can be held in the focus of attention increases over development (e.g., Cowan, Saults, & Elliott, 2002). Through a series of manip- ulations and controls, Cowan and colleagues have shown that other cognitive changes, like strategies or knowledge, cannot account for developmental improvements (see Cowan, 2016, for review). Thus, Cowan (2013) proposed that such improvements reflect increases in capacity itself, although he noted that the underlying source of capacity increases was unknown.

One proposed explanation of capacity increases came from studies assessing multifeature object memory. Children (7- to 10-years-old) and adults showed comparable performance on change detection trials requiring memory for one versus two features per object, which was interpreted as evidence for “inte- grated object” representations (Riggs, Simpson, & Potts, 2011; Vogel, Woodman, & Luck, 2001; but see below for conflicting evidence). Vogel, Woodman, and Luck (2001) and Riggs, Simp- son, and Potts (2011) cited a neural synchrony model proposed by Raffone and Wolters (2001), in which different neural populations represent different features (e.g., color vs. orientation), and fea- tures from the same object fire synchronously. Capacity limits come from the temporal resolution of synchronous firing: as the number of objects increases, differentiating the timing of different cell assemblies that represent each object becomes more difficult. Riggs et al. (2011) extended this model to account for develop- mental change, proposing that the temporal resolution of synchro- nous firing improves over development.

As an alternative to slot-like explanations, Bays and Husain (2008) characterized VWM as a continuous pool of resources that could be divided among an unlimited number of representations. From this view, adults’ change detection performance declines as the number of items increases because representations decrease in precision (see Suchow et al., 2014, for review). Evidence for this view came from a delayed estimation task (Wilken & Ma, 2004) in which participants recalled an item’s feature value (i.e., clicking on a continuous color wheel, or adjusting the orientation of a response bar, to match the remembered color or orientation) rather than judging “same” or “different” as in change detection. Pooling responses across many trials provides an estimate of the fidelity of memory representations, and results indicate that representational precision decreases as the number of items in memory increases (e.g., Bays & Husain, 2008).

Three studies have assessed developmental changes in precision with this paradigm. Burnett Heyes and colleagues found that 7- to

T hi

s do

cu m

en t

is co

py ri

gh te

d by

th e

A m

er ic

an Ps

yc ho

lo gi

ca l

A ss

oc ia

tio n

or on

e of

its al

lie d

pu bl

is he

rs .

T hi

s ar

tic le

is in

te nd

ed so

le ly

fo r

th e

pe rs

on al

us e

of th

e in

di vi

du al

us er

an d

is no

t to

be di

ss em

in at

ed br

oa dl

y.

1475REAL-TIME STABILITY IN FEATURE BINDING

 

 

13-year-old children showed the same load-precision trade off as adults, and with better precision in older children (Burnett Heyes, Zokaei, van der Staaij, Bays, & Husain, 2012), and that precision improved with age when testing the same children 2 years later (Burnett Heyes, Zokaei, & Husain, 2016). Probabilistic modeling of responses suggested that age-related improvements related to decreased noise in memory representations. In contrast, however, a study comparing younger (7–9 years) versus older (10–12 years) groups of children found no age-related change in precision, but rather a decrease in incorrect-target responses (i.e., recalling a feature from an uncued item; Sarigiannidis, Crickmore, & Astle, 2016). Together these results indicate developmental improve- ments in children’s performance, but conflicting analyses make it unclear whether they reflect increasing precision, better selection from memory, or both.

Across these domains of research, behavioral evidence suggests that VWM improves in multiple ways from infancy through mid- dle childhood. During infancy, memory representations are built more quickly, are maintained and used more robustly, and increase in complexity with development. These changes have generally been attributed to brain maturation. During childhood, both capac- ity and precision of memory increase, but theories have been designed to address only one of these characteristics. As a way to bridge across these tasks and age groups to provide a more com- prehensive theory, Simmering (2016) proposed a dynamic systems approach to understanding how VWM functions and develops, which we describe next.

The Cognitive Dynamics Theory of Visual Working Memory Development

The cognitive dynamics theory of VWM development is a dynamic systems approach proposed to bridge from infant devel- opment to adulthood (Simmering, 2016). Prior theories have been disconnected largely due to the different types of tasks used to assess VWM in infants, children, and adults. Dynamic systems approaches are well-suited to reconciling such task differences through their foundational concepts. In particular, dynamic sys- tems theories conceptualize cognition and behavior as part of a larger system including endogenous and exogenous factors, with the same contributions driving change across timescales (Fogel & Thelen, 1987; Smith & Thelen, 2003). Importantly, no single component of the system has priority in explaining behavior and development, meaning the structure of task is just as important to consider as the structure of the cognitive system. Historically, theories of cognition and development sought to devise behavioral tasks that tap into particular cognitive constructs, without consid- ering how the task itself creates the behavior it measures (see Smith, Thelen, Titzer, & McLin, 1999, for an illustration using the A-not-B paradigm).

In the context of VWM development, a dynamic systems ap- proach can unite results from looking paradigms in infancy with tasks designed to assess capacity in children and adults. The cognitive dynamics theory specifically emphasizes the continuity and interdependence of processing within these different tasks (Simmering, 2016), as opposed to classic information processing descriptions of memory systems (e.g., Atkinson & Shiffrin, 1968) that posit separable stages of attention, encoding, storage, and retrieval. In the cognitive dynamics theory, rather than considering

a separate encoding process that precedes storage, the same tem- porally continuous processes form, maintain, and use memory representations in service of behavior (Johnson et al., 2014; Sim- mering, 2016).

The cognitive dynamics theory has been formalized into a computational model to illustrate the explanation of behavior across tasks and development. Computational modeling is a valu- able approach to understanding cognitive processes, but is under- represented in developmental theories (Simmering, Triesch, Deàk, & Spencer, 2010). Computational approaches face a number of challenges in reaching a broad developmental audience, as they must be constructed to address the specific characteristics of the task(s) of interest. This specificity can be a double-edged sword: It allows for incremental extensions to predict closely related behav- iors (e.g., Simmering & Patterson, 2012), but requires substantial modification to encompass more distant tasks and phenomena (see Simmering, 2016, for discussion). Thus, to be most effective, computational approaches must balance specificity and generality to advance our understanding of cognitive and developmental processes (Simmering & Spencer, 2008).

The current investigation strikes this balance by using models as a theoretical framework, showing general predictions without spe- cific implementation of the tasks. The cognitive dynamics theory of VWM development is instantiated in a dynamic neural field architecture, in which features are represented in continuous neural fields along metrically specific dimensions (i.e., location, color, orientation). Nodes within these fields are connected such that local excitation supports activation among similarly tuned nodes, which allows for localized “peaks” of activation to form as repre- sentations of features. The architecture to simulate VWM tasks includes two excitatory layers coupled to a shared inhibitory layer (Johnson & Simmering, 2015). The excitatory layers simulate perceptual processing versus working memory representation through different strength of excitatory and inhibitory connections. In particular, weaker connectivity in the perceptual layer leads to input-driven representations: Activation remains above threshold only in the presence of input, reflecting perceptual processing of visual information. Stronger connectivity in the working memory layer produces self-sustaining representations: Peaks maintain above-threshold activation after input is removed, reflecting mem- ory for prior visual information. Importantly, activation in these layers interacts continuously throughout the task, with excitation projecting from the perceptual layer to the inhibitory and working memory layers, as well as from working memory to the inhibitory layer, with inhibition projecting back into both excitatory layers (Johnson & Simmering, 2015). Through the balance between excitation and inhibition, items compete for representation, result- ing in a limited capacity (cf. Franconeri, Alvarez, & Cavanagh, 2013). This is not a hard limit as in slot-like conceptualizations, but rather varies according to task demands (see Johnson et al., 2014, for illustrative simulations).

These interactions among layers provide an implicit mechanism for comparison between items held in memory and new inputs. Specifically, items (peaks) in the working memory layer are main- tained through local excitation (within the layer) and lateral inhi- bition (from the inhibitory layer). Because the inhibitory layer also projects to the perceptual layer, activation at the corresponding values along the metric dimension (i.e., features similar to those held in working memory) is relatively suppressed. Conceptually,

T hi

s do

cu m

en t

is co

py ri

gh te

d by

th e

A m

er ic

an Ps

yc ho

lo gi

ca l

A ss

oc ia

tio n

or on

e of

its al

lie d

pu bl

is he

rs .

T hi

s ar

tic le

is in

te nd

ed so

le ly

fo r

th e

pe rs

on al

us e

of th

e in

di vi

du al

us er

an d

is no

t to

be di

ss em

in at

ed br

oa dl

y.

1476 SIMMERING AND WOOD

 

 

this reflects reduced perceptual processing of familiar items. By contrast, novel items (i.e., feature values not held in working memory) produce a novelty signal through strong activation in the perceptual layer. When integrated with a fixation system, this mechanism of comparison can account for habituation (Perone & Spencer, 2013b) and novelty preferences (e.g., Perone & Spencer, 2014) in looking paradigms. When coupled to a “same”/“different” response system, this same mechanism can explain children’s and adults performance in the change detection task (e.g., Simmering, 2016) and a single-item color discrimination task (Simmering & Patterson, 2012). Thus, the dynamic model architecture specifies the real-time processes of encoding, maintenance, and comparison required to recognize familiarity and detect novelty across VWM tasks.

The cognitive dynamics theory incorporates a second key con- cept from dynamics systems in the integration across timescales (Fogel & Thelen, 1987; Smith & Thelen, 2003). From this per- spective, understanding how behavior emerges in the moment of a task can provide insight into how behavior changes over develop- ment, as both reflect the interacting components of the system. Specifying the processes that support behavior across tasks in a model can provide a test of potential mechanisms to explain development (see Simmering & Schutte, 2015, for discussion). A single type of developmental change in the dynamic model— strengthening connectivity within and between layers—can ac- count for improvements in habituation (Perone & Spencer, 2013b), visual paired comparison (Perone et al., 2011; Perone & Spencer, 2013a, 2014), change detection (Simmering, 2016; Simmering et al., 2015), and color discrimination (Simmering & Patterson, 2012), plus a range of spatial memory tasks (Simmering, Schutte, & Spencer, 2008). This change in connectivity formalizes another central concept from dynamic systems theory: dynamic stability (which we term “real-time” stability to contrast with long-term stability).

The concept of real-time stability is most easily illustrated through motor development, for example, observing an infant learning to reach or walk. Early forms of behavior are uncoordi- nated, unpredictable, unreliable—collectively, unstable. Through repetition in the coordination of the processes supporting the behaviors, they become more predictable and reliable over time, more broadly applied (e.g., grasping objects of different sizes, walking in different shoes), and more resistance to disturbance (e.g., not dropping an object if something contacts the hand, not falling on uneven surfaces). Simmering (2016) extended this no- tion to VWM, conceptualizing real-time stability as the collective improvement in the formation, maintenance, and use of represen- tations in service of behavior. Importantly, the continuous process- ing instantiated in the dynamic model show how increases in real-time stability could arise through strengthening neural con- nectivity (see Simmering, 2016, for details).

The cognitive dynamics theory was proposed to bridge research across tasks and age groups to provide a more comprehensive account of how VWM functions and develops (Simmering, 2016). This theory and model formalization have generated specific pre- dictions across a range of visuospatial tasks over development (see Simmering & Schutte, 2015, for review). The current article tests the real-time stability hypothesis further in the context of feature binding in VWM. Specifically, we test the prediction that unstable memory early in development leads to poor localization of fea-

tures, which impairs feature binding in VWM. In the next section, we review the major theories of feature binding to provide further context to our predictions.

Feature Binding in Visual Working Memory

There are three notions of binding in visual cognition. The first addresses how features from different dimensions (e.g., color and shape) may combine in object representations (e.g., Treisman & Gelade, 1980). The second concerns how features are bound to locations, such as the relative locations of two colors on an object (e.g., Dessalegn & Landau, 2008) or multiple colored squares within a change detection display (e.g., Cowan, Naveh-Benjamin, Kilb, & Saults, 2006). Third is contextual binding, in which object representations are linked to the learning context (e.g., Holling- worth, 2007). We investigated the first type of feature binding, and therefore limit our review to related studies, although it is possible that all three rely on similar processes.

"Is this question part of your assignment? We can help"

ORDER NOW