Word-Object Learning via Visual Exploration in Space (WOLVES): A Process Model of Cross-Situational Word Learning

• In Event: It’s about time: Integrating the multiple timescales of attention and memory in early word learning.

Abstract

Infants, children and adults have been shown to track co-occurrence across ambiguous naming situations to infer the referents of new words. The extensive literature on this cross-situational word learning (CSWL) ability has produced support for two theoretical accounts—associative learning (AL) and hypothesis testing (HT)—but no comprehensive model of the behaviour. Here, we present an overview of WOLVES, a neural process account of CSWL grounded in real-time psychological processes of memory and attention that explicitly models the dynamics of looking at a moment-to-moment scale and learning across trials.

Figure 1 shows a schematic of WOLVES. The model integrates the word-object learning (WOL) model shown in green (Samuelson, Smith, Perry, & Spencer, 2011; Samuelson, Spencer, & Jenkins., 2013) with a model of visual exploration in space (VES) shown in red (Schneegans, Spencer, & Schöner, 2016). These two models share the common elements in the overlapping shaded boxes (aspects of spatial working memory and a scene representation). Note that the VES model is also an integrative model in its own right, bringing together earlier models of the neural processes that operate in early visual processing, a model of visual working memory, and a model of spatial working memory. These models are integrated in a way that is consistent with neural evidence for dorsal (‘where’ or ‘how’) and ventral (‘what’) pathways in the brain.

We will show that WOLVES quantitatively captures data from 12 studies of CSWL with adults and children (accurately simulating 177 data values), thereby providing a comprehensive account of data purported to support both AL and HT accounts. Direct model comparison with two other models proposed in the literature (Kachergis et al., 2012, 2013, 2017 and Pursuit by Stevens et al, 2017) shows that WOLVES outperforms these competitor models.

Critically, we offer the first developmental account of CSWL, providing insights into how memory processes change from infancy through adulthood. We will show that WOLVES captures developmental changes in CSWL behaviours via manipulations to only two parameters: tau_Build, which specifies the timescale of memory trace formation, and tau_Decay, which specifies the timescale of memory-trace forgetting. In particular, over development we decreased tau_Build – accelerating memory formation – and increased tau_Decay, making the memory trace dynamics more resistant to decay. These were the only parameter changes required to quantitatively capture results from 7 developmental studies.

Conclusions from this modelling work are numerous. WOLVES shows that visual exploration and selective attention in CSWL are both dependent on and indicative of learning. Further, learning is driven by real-time synchrony of words and gaze-fixations and constrained by memory processes operating over multiple timescales. Additionally, WOLVES explains how performance is impacted by the structure of test paradigms, how within- and across-trial competition produce mutual exclusivity, and how previously observed individual differences can emerge from learning in the task. Finally, WOLVES offers a framework for making developmental predictions about how CSWL abilities should change longitudinally and how individual differences in working memory and attentional processes should be related to early word learning.

Authors

• Larissa K Samuelson, University of East Anglia
  Presenting Author

• Ajaz Bhat
  Non-Presenting Author

• John P Spencer, University of East Anglia
  Non-Presenting Author