Connectionist modeling of developmental changes in infancy: Approaches, challenges, and contributions.

Abstract

Connectionist models have been applied to many phenomena in infant development including perseveration, language learning, categorization, and causal perception. In this article, we discuss the benefits of connectionist networks for the advancement of theories of early development. In particular, connectionist models contribute novel testable predictions, instantiate the theorized mechanism of change, and create a unifying framework for understanding infant learning and development. We relate these benefits to the 2 primary approaches used in connectionist models of infant development. The first approach employs changes in neural processing as the basis for developmental changes, and the second employs changes in infants' experiences. The review sheds light on the unique hurdles faced by each approach as well as the challenges and solutions related to both, particularly with respect to the identification of critical model components, parameter specification, availability of empirical data, and model comparison. Finally, we discuss the future of modeling work as it relates to the study of development. We propose that connectionist networks stand to make a powerful contribution to the generation and revision of theories of early child development. Furthermore, insights from connectionist models of early development can improve the understanding of developmental changes throughout the life span. Keywords: connectionist modeling, neural network modeling, parallel distributed processing, infant developmentSince the groundbreaking work of Rumelhart and McClelland in the 1980s, there has been an increasing interest in and research on the application of connectionist models to early human development. Connectionist models are instantiations of theories about the mechanisms that underpin particular behaviors. Building computational models allows for the exploration of the interaction of numerous factors both internal and external to the organism that typically contribute to a behavior, which often can be too complex to specify through verbal theory alone . These computational models provide researchers with a number of ways to explicitly test theoretical assumptions and develop novel and testable predictions.However, in our view, for many developmental scientists the contribution of models to an integrated understanding of development is far from clear. Although a given model might provide output that is similar to the behavior of infants, it often remains to be seen whether the model's results and the behavioral results occur for the same reasons. For example, a network may simulate effectively infants' ability to discriminate between two objects that differ along multiple features, but it is possible that the features used by the network for discrimination are different from those used by infants. Computational models have also been criticized for using overly technical terms and notations that may be offputting to nonexperts (Klahr, 2004); for not being explicit about the source of their starting states (Oakes, Newcom...

Statistical learning of parts and wholes: A neural network approach.

“…Simulation 2: Fiser and Aslin (2005) Experiment 1 Although Giroux and Rey's (2009) results can be explained by a standard SRN trained on syllable prediction, this shouldn't be taken to imply that such a model is the most appropriate way to account for language learning more generally. SRNs are often viewed as a specific type of neural network that is distinct from other types that have been applied to developmental data (see Yermolayeva & Rakison, 2014), and some researchers consider the application of SRNs to tasks other than prediction to be "a fundamental change in the way in which SRNs are conceptualized" (French et al, 2011, p. 621). We think these views are too narrow.…”

What can we learn from learning models about sensitivity to letter-order in visual word recognition?

“…Consequently, the present approach can, in principle, trace clear developmental predictions regarding how reading and its supporting representations are shaped in a given language through exposure to larger and larger corpora of words. In turn, this type of approach can capture how readers gradually derive the statistical properties of their language, thereby determining their reading behaviour (see Yermolayeva & Rakison, 2013, for additional discussion related to the connectionist modeling of development).…”

Perception of an object’s global shape is best described by a model of skeletal structure in human infants

“…More specifically, because habituation/dishabituation can be conceived as a measure of alignment between the stimulus and the infant’s internal representation ( Mareschal et al, 2000 ; Westermann and Mareschal, 2004 ), we tested models by feeding their outputs into an autoencoder and measuring the error signal across habituation and test phases (see Methods). Like habituation paradigms, the error signal of an autoencoder reflects the degree of alignment between the internal representation of the model and the input stimulus (for review, see Yermolayeva and Rakison, 2014 ). Unlike conventional classifiers, which often require multiple labeled contrasting examples (e.g.…”