This is a slightly extended version of a paper that appeared in: Proc. of the 8th Neural Computation Workshop (NCPW8), 28-30 August 2003, Univ. of Kent, Canterbury, England, UK, H. Bowman and C. Labiouse (eds.), Progress in Neural Processing, World Scientific, 2003. A computational model of the emergence of gaze following Eric Carlson and Jochen Triesch Cognitive Science Department, UC San Diego, 9500 Gilman Drive, La Jolla, CA 92093-0515, USA E-mail: {eccarlson,triesch}@ucsd.edu Abstract We present a computational model of the emergence of gaze following in infant caregiver interac- tions. Using the model we explore the plausibility of the hypothesis that gaze following is a skill that infants acquire because they learn that monitoring their caregiver’s direction of gaze allows them to predict where interesting objects or events in their environment are. In particular, we demonstrate that a specific basic set of mechanisms is sufficient for gaze following to emerge and we show how plausible alterations of model parameters motivated by findings on developmental disorders lead to impairments in the learning of gaze following. 1 Introduction The capacity for shared attention is a cornerstone of human social intelligence. It is an early foundational skill of vital importance for learning in social contexts. By the end of their first year of life, normally developing infants show robust gaze following[4] as well as a variety of other shared attention skills, such as pointing and requesting behaviors. How do these skills develop? Why is their emergence disrupted in certain developmental disorders? Our particular focus is the emergence of gaze following behavior. We propose a computational model for the emergence of gaze following based on reinforcement learning. The model demonstrates that gaze following behavior can be robustly learned with a generic basic set of mechanisms[6]. Our learning-based view is in stark contrast to earlier nativist/modularist accounts of the emergence of shared attention skills[1, 12] and offers very different explanations for failures of the emergence of gaze following in developmental disorders such as autism and Williams syndrome. The work described in here is part of the MESA project at UC San Diego (Modeling the Emergence of Shared Attention, http://mesa.ucsd.edu), a larger effort to understand the emer- gence of shared attention in normal and abnormal development through closely integrating observa- tional studies with computational modeling approaches[6]. 2 The Basic Set Hypothesis We start from the premise that earlier nativist/modularist theories of the emergence of shared attention skills[1] are of little explanatory power. The problems of such theories have been eloquently pointed out by many researchers[15, 5] and will not be discussed here. We believe that better theories will be based on a careful analysis of the developing infant’s interactions with her environment during development. For the case of of gaze following, Moore has suggested that gaze following may emerge because infants learn that the caregiver’s direction of gaze is predictive of where interesting things are in the environment[10]. This idea is at the core of our model. Our goal is to formalize this hypothesis into a concrete mathematical model and to test its plausibility and implications using computer simulations. Previously, we hypothesized that gaze following may emerge from an interplay of a basic set of mechanisms[6] which may contain the following: • a set of perceptual and motivational biases and habituation mechanisms driving the infant to look at and shift attention between “interesting” visual stimuli 1