Talk

• Date: Friday, July 15th, 2005.
• Time: 14:00-15:20
• Venue: Multimedia Conference Room (next to TV Conf Room)
• Speaker: Christopher G. Prince (Department of Computer Science, University of Minnesota Duluth)
• Title: Towards robotic models of infants: Ongoing emergence and audio-visual sensory intergration
• Abstract:
  We have been following a three-part approach to epigenetic robotics. First, we are formulating a theory related to the developmental growth by robots of various psychological abilities. Second, we are constructing sensory-oriented computational models of infants audio-visual integration skills. Third, we are conducting empirical research with infants regarding audio-visual integration. Our vision is to gradually formulate this theory in terms realizable in robots, to gradually build-up our robot systems to reach towards the theory, and throughout, to incorporate comparisons and findings from developmental psychology. The theory focuses on the idea of ongoing emergence (Prince et al., to appear; see also Prince, 2001), which refers to the continuous development, integration, and incorporation of new skills. Ongoing emergence commonly occurs in human infants (e.g., in walking, word learning, and visual object skills) and differs markedly from accomplishments to date in epigenetic robotics. In particular, the continuous acquisition of robotic skills, and the general incorporation of new robotic skills with existing robotic skills in the repertoire of an individual robot are not well established. Our sensory-oriented computational modeling research follows from the observation that young infants learn better in certain situations involving contingent stimuli. For example, 7-month-old infants can learn word-object mappings better when provided with synchronized speech-visual signals (Gogate & Bahrick, 1998, 2001). In this regard, we have been constructing models of infant audio-visual synchrony detection (Prince & Hollich, 2005; Prince et al., 2004). These models qualitatively and quantitatively detect audio-visual synchrony based on an algorithm computing Gaussian mutual information between two input channels (Hershey & Movellan, 2000). While the tasks performed by model and infant are not yet the same, the model compares favorably to aspects of the published infant literature. In a new study, we have also found that 7-8 month-old infants, who looked at face motion from two people but heard speech audio from only one, have moment-by-moment looking behavior that correlates significantly even with this basic model (r = .30, p < .001; Hollich et al, 2004). Our present research goals include creating a framework for sensory-oriented modeling to enable more flexible creation and use of models, using this framework to create models of infants audio- visual learning based on synchrony detection, and determining how contingency detection may contribute to processes of ongoing emergence.