Modeling U Shaped Performance Curves in Ongoing Development

Modeling U Shaped Performance Curves in Ongoing Development Anthony F. Morse (anthony.morse@plymouth.ac.uk) Tony Belpaeme (tony.belpaeme@plymouth.ac.uk) Angelo Cangelosi (angelo.cangelosi@plymouth.ac.uk) Centre for Robotics and Neural Systems, University of Plymouth, Devon, PL4 8AA, UK Caroline Floccia (caroline.floccia@plymouth.ac.uk) School of Psychology, University of Plymouth, Devon, PL4 8AA, UK Abstract This paper details a simple and general account, and model, of the U-shaped curve phenomena apparent in many developmental psychology experiments. The model replicates both the general form of the U-shape performance in ongoing development and accounts for additional observations in the psychology literature such as the effect of noise in Switch task experiments. This leads to predictions both in psychology and neuroscience, and establishes an alternative hypothesis, which is simpler, more detailed, more predictive, and more general than those already established in the literature. This approach is also suitable for embodied robotic modeling of development. Keywords: Cognitive modeling; neural networks; epigenetic robotics; language acquisition; development; U-shaped curve, Self-Organizing Maps; Active Hebbian learning. Ongoing Development in Humans and Robots This paper presents a novel neuro-computational approach to modeling cognitive development, in particular for the investigation of the phenomenon of U-shaped performance curves in development. The model is based on refinements of the associative learning mechanism recently proposed as part of the Epigenetic Robotics Architecture (ERA) (Morse, DeGreeff, Belpeame, & Cangelosi, 2010): a neural cognitive architecture for general, scalable and embodied learning and modeling of psychological function. This architecture is particularly suited to model the role of embodiment and agent-environment interaction in development. Modeling even a part of the process of development itself is an inherently general proposition, as humans we all go through significant physical and mental development from conception into adulthood and old age. Some of this development can be attributed to physical growth or other factors principally under genetic control. This is the case, for example, of the development of the musculoskeletal system for walking, or the mental and physical effects of puberty. Other developmental transitions are more obviously influenced by our physical and social environments, such as learning to read, or which languages you speak. But no single developmental phenomenon results wholly from nature or nurture alone (Karmiloff-Smith, 2000; Oyama, 2000a, 2000b). We are not static agents untouched by our past and we are more than the unfolding of our genetic program. The environment always plays a role, as we shall see in the experiments herein. Recognizing this, and in contrast to a growing body of modeling work in which adaptation does not occur during the lifetime of an agent (e.g. artificial evolution), is the field of Epigenetic or Developmental Robotics (Metta & Berthouze, 2006, p. 129). While there are clear technological outcomes from endowing robots with the capacity to learn and develop, herein we focus our modeling efforts to aid and refine our understanding of human development. As the general model of U-shape learning proposed here is based on the Epigenetic Robotics Architecture (already used in development robotics experiments, Morse, Belpaeme, Cangelosi, & Smith, 2010; Morse, DeGreeff, et al., 2010), the extension of this study to new robotics experiments is facilitated. So what is (ongoing) development in humans? From Experimental Psychology we know that much of development is not simply the linear acquisition of new skills / abilities / knowledge. Instead, the outward effects of development often happen in non-linear stage-like transitions, and rarely is it the case that some new behavior or ability is simply added to an otherwise unchanged pile. A commonly found phenomenon in developmental psychology is known as the U-shaped curve; here previously stable abilities become temporarily absent or disrupted for a period of time (sometimes months) before returning in a changed but stable form as new competencies emerge. This U-shaped pattern of behavior reoccurs again and again throughout the child development literature and is not specific to the involvement of any particular modality or physicality. This is, of course, not the only pattern of development to be found but its frequent occurrence combined with independence from any particular mode of expression or sensory modality would seem to indicate a common feature of the learning systems involved. As such, competing accounts of this U-shaped pattern of behavior can potentially have far-reaching impact on the cognitive sciences. U-shaped curve phenomena appear to be independent from any particular task or modality as the following prominent examples demonstrate: For example, Bosch and Sebastian-Galles argue that initially, bilingual infants track statistical regularities across the two languages, leading to their temporary inability to discriminate acoustically similar