Abstract
In this chapter, the authors explore social constructivist theories of emotion, which suggest that emotional behaviors are developed through experience, rather than innate. The authors' approach to artificial emotions follows this paradigm, stemming from a relatively young field called developmental or ‘epigenetic' robotics. The chapter describes the design and implementation of a robot called MEI (multimodal emotional intelligence) with an emotion development system. MEI synchronizes to humans through voice and movement dynamics, based on mirror mechanism-like entrainment. Via typical caregiver interactions, MEI associates these dynamics with its physical feeling, e.g. distress (low battery or excessive motor heat) or flourishing (homeostasis). Our experimental results show that emotion clusters developed through robot-directed motherese (“baby talk”) are similar to adult happiness and sadness, giving evidence to constructivist theories.
Top1. Introduction
Some of the most revolutionary ideas in brain science are coming from cribs and nurseries. – Patricia Kuhl
Are emotions innate? Recently, the popular Darwinian theory that basic emotions – such as happiness, sadness and anger – are “hard-wired” through evolution has been called into question. Psychologists have collected growing evidence (see reviews, (Mason & Capitanio, 2012; Camras & Shuster, 2013; Barrett, 2006)) that emotions may not in fact be completely a product of innate biology. Instead, “social constructivist” theories (Averill, 1980) point to experiences and environment as a prime factor for the development of emotions:
While there is little doubt that what we call fear, anger, and sadness refer to real (i.e., observable) phenomena and important parts of human experience, the weight of scientific opinion appears to be shifting away from the view that a few specific emotions are natural and universal kinds, laid down in the biology of humans and other animals (nature), in favor of a larger place for experience (nurture) in all emotions– (Mason & Capitanio, 2012, p. 239)
Let us briefly illustrate this view with evidence from infant psychology, animal behavior and cultural emotion psychology.
In infant developmental psychology, (Camras L. A., 2011) has pointed out several phenomena that support the constructivist view. Firstly, emotional facial expressions were observed in infants where the emotion was not expected to occur: 5-7 month olds showed prototypical surprise expressions while bringing familiar objects into their months (Camras, Lambrect, & Michel, 1996). Secondly, emotional expressions were not observed in contexts during which the emotion should have occurred: 10-12 month olds in the visual cliff procedure rarely produced the fear expression, even though their other behaviors showed that they did in fact experience fear (Hiatt, Campos, & Emde, 1979). Finally, Camras and her colleagues also found that negative emotion classes (such as anger, sadness and fear) did not seem to differentiate well in infants as old as 11 months, suggesting that they all corresponded to a general negative “distress” affect. (Camras L. A., 2011).
Studies on atypical caregiver conditions in young animals also support the social constructivist view. In a study with rats, it was shown that when mother rats’ maternal style contained more licking and grooming, their pups grew up to be less fearful, with decreased hormonal reactions to stressors (Kaffman & Meaney, 2007). In studies with rhesus monkeys, maternal separation early in life affected gene expression in brain regions controlling socio-emotional behaviors, with a correlation on the timing of the separation (Sabatini, Ebert, Lewis, Levitt, Cameron, & Mirnics, 2007). For example, monkeys separated from their mother at 1 week of age showed less expression of the gene GUCY1A3 (associated with social-seeking comfort behaviors), compared to the 1 month old separation condition.
Human studies on atypical early caregiving conditions also exist, though are more rare due to ethical issues. For instance, observations on postinstitutionalized (PI) children, such as those adopted from Eastern European orphanages after World War II, provide evidence that nurture is important for emotional intelligence. According to (Fries & Pollak, 2004), PI individuals had difficulty matching appropriate faces to happy, sad and fearful scenarios, yet were able to match angry faces just as well as controls. We refer the reader to further observations on the effect of early adverse rearing in work by Tottenham et al., e.g., (2011). In addition, temperament, while thought to have an inherent basis from birth, is not stable over the lifetime. In a short-term longitudinal study, (Calkins, 2002) found that infants who experienced negative parenting continued to show high anger/frustration levels, though it was not the case for infants who experienced positive parenting. Studies such as these emphasize the importance of emotional input early in life.
Key Terms in this Chapter
Motherese: A simplified type of speech spontaneously spoken by caregivers to infants. Typically it contains exaggerated intonation and rhythm, a higher pitch, and more pronounced variations compared to normal speech. Also known as baby talk or infant-directed speech (IDS).
SIRE Paradigm: A paradigm using speed, intensity, regularity and extent (SIRE) to represent an emotion across modalities. For instance, sadness has been linked to slow, low intensity, regular and small dynamics in movement, as well as in voice and music.
Gut or Physical Feeling: The state of physical flourishing (homeostasis) or distress (out of homeostasis) in an individual.
Gaussian Mixture Model (GMM): A probabilistic model used to represent data as a mixture of normal distributions. It is commonly used for unsupervised learning and clustering, which means that clusters can be created without labels. It is similar to k-means clustering, except that when used for recognition, it outputs the probability that a new data point belongs to a cluster, not a binary value.
Bayesian Information Criterion (BIC): A criterion for selecting the best model given a dataset. It penalizes models that use too many variables to explain the data (overfitting). A lower BIC implies 1) a better fit, and/or 2) fewer explanatory variables.
Developmental Robotics or Epigenetic Robotics: A relatively new scientific field that aims to study the developmental mechanisms and architectures for lifelong learning in machines. Typically it involves formalizing, validating and extending models from neuroscience, developmental psychology, and evolutionary biology, specifically by attempting to implement the models in robots. Results are expected to feedback into existing theories, or produce novel theories about human and animal development.
Entrainment: The synchronization of organism to a rhythm usually produced by another social actor. Humans can entrain to the beat, for instance, by dancing or tapping their foot, and fireflies are also known to flash in synchrony. In this chapter, we refer to the entrainment in speed, intensity, regularity and extent between the voices and movements of two agents.
Social Constructivist Theory: A theory that an individual’s learning is constructed through interaction with others in a group. It suggests that cognitive development is influenced by culture and social context.
Multimodal Emotional Intelligence (MEI): A robot system with the ability to understand, represent and express emotions in multiple modalities, such as voice, movement, gait or music.