Running head: SURPRISE AS VICARIOUS PREDICTION ERROR Surprisingly unsurprising! Infants’ looking time to improbable events is modulated by others’ expressions of surprise In preparation Yang Wu Hyowon Gweon Department of Psychology, Stanford University Keywords: cognitive development; social cognition; surprise; prediction error; emotion, statistical inference; looking time Address for correspondence: Corresponding Author Name: Yang Wu Department of Psychology Stanford University 450 Serra Mall, Jordan Hall 280 Stanford, CA 94305 E-mail: [email protected] 1 Abstract Research in diverse disciplines suggests that agents’ own prediction error enhances their learning. Yet, human learners also possess powerful capacities to learn from others. Here we ask whether infants can use others’ expressions of surprise as vicarious prediction error signals to infer hidden states of the world. First, we conceptually replicated Xu & Garcia (2008), showing that infants (12.0-17.9 months) looked longer at improbable than probable sampling outcomes (Experiment 1). Then we added emotional cues to the design (Experiment 2). Before revealing an outcome to an infant, the experimenter looked at the outcome and showed either an unsurprised or surprised emotional expression. While infants still looked longer at the improbable than the probable outcome following the experimenter’s unsurprised emotional expression, this trend was reversed following the experimenter’s surprised expression. Thus, although for decades infants’ looking time has been assumed to be longer for events that are unexpected or surprising, these results suggest that infants’ inferential abilities in social contexts are powerful enough to make surprising events “surprisingly unsurprising.” 2 Human learners accomplish remarkable achievements in their first years of life – understanding minds, numbers, and languages, to name a few (Carey, 2009; Gopnik, 2012). How children learn has inspired research in diverse scientific domains, and has been intensively studied for several decades. One challenge for learning, noted by this literature, is that our learning environment is often complex, dynamic and full of information. To make learning efficient, or even possible at all, learners must understand when to attend to some information, when to ignore it, and from whom to learn. One of the most powerful proposals for how learning occurs states that agents identify opportunities for learning based on prediction error (e.g., Kulkarni, Narasimhan, Saeedi, & Tenenbaum, 2016; Schultz, 2016; Silver et al., 2018). When there is a discrepancy between what an agent expects to happen and the actual outcome, the agent uses the discrepancy to modify its behavior. The neural and computational bases of such error-driven learning have been extensively studied in both animal models and human adults (see Schultz, 2016 for a review). This learning mechanism has also been a central notion of theoretical models in various domains, including reinforcement learning, perceptual inference, cognition, and decision-making (see Den Ouden, Kok, & De Lange, 2012). This type of error-driven, surprise-induced learning has also been found in human infants and children. When infants and children observe an event that is inconsistent with their predictions or expectations, they show a range of behavioral, neural and physiological changes. The most well-known behavior is that infants look longer at an event that violates their expectation than an event that does not. Accordingly, infants’ looking time to expected vs. unexpected events has been used in developmental research as a popular dependent measure (i.e., violation-of-expectation paradigm). Infants also show changes in pupil dilation (Jackson & Sirois, 2009), cerebral blood flow (Wilcox, Bortfeld, Woods, Wruck, & Boas, 2005) and brain electrical activity (Berger, Tzur, & Posner, 2006) following unexpected events. Critically, infants and older children also explore more when events are unexpected, and show enhanced 3 learning following surprising events (e.g., Bonawitz, van Schijndel, Friel, & Schulz, 2012; Chandler & Lalonde, 1994; Phelps & Woolley, 1994; Schulz, 2012; Stahl & Feigenson, 2018). Nonetheless, we often miss out on a lot of surprising events. Sometimes we fail to pay attention or lack visual access to those events. Critically, when one lacks the relevant world knowledge, one might fail to find an event surprising even when it is in principle surprising; for instance, if you know little about the weather in Lima (where it almost never rains), you won’t be surprised when it rains in Lima. Young learners, in particular, may often fail to detect surprising events. Despite their sophisticated, early- emerging knowledge of the physical and social world (e.g., core principles of physical objects; Spelke, Breinlinger, Macomber, & Jacobson, 1992), events that violate this knowledge are quite rare (e.g., a car floating in the air), making many potential opportunities for learning go unnoticed. Thus, although surprise-induced learning may be one fundamental mechanism that explains how agents learn broadly, it does not explain some critical aspects of human learning. However, learning does not occur in isolation; young learners benefit from others who already know much about the world (Csibra & Gergely, 2009; Boyd & Richerson, 1988). Here we propose that young children may not only learn from their own prediction error, but also use others’ expressions of surprise as vicarious prediction error signals to identify special opportunities for learning. Such cues are abundant in children’s everyday life: Parents, caregivers, and educators do not only express various emotional expressions when they interact with young children, but also tend to exaggerate them for pedagogical and communicative purposes. In particular, expressions of surprise are often used to attract children’s attention to an interesting (and potentially informative) event even when it may not necessarily be surprising for adults. This proposal is supported by the literature on early emotion understanding. From infancy on, children show remarkably early-developing sensitivity to others’ emotional expressions, and even represent others’ emotions as caused by events in the world and others’ goals, desires, and beliefs (see 4 Reschke, Walle, & Dukes, 2017 and Wellman, 2014 for reviews). Infants and children can also perform backward inferences, using observed emotional expressions to reason backward about their unobserved causes in the environment (Wu, Muentener, & Schulz, 2017) and hidden mental states of other people (Wu & Schulz, 2018). These studies suggest that even young children have a sophisticated understanding of others’ emotional expressions, and use those emotional expressions as powerful sources of information to guide their understanding of the world. Research on the emotional expression of surprise, however, has primarily studied this emotional expression as an affective consequence of one’s false beliefs (e.g., Hadwin & Perner, 1991; Ruffman & Keenan, 1996; Harris, 1989; Wellman & Banerjee, 1991; Scott, 2017). Additionally, a recent study found that 7-year-olds expect others to be surprised in response to a low-probability event (Doan, Friedman, & Denison, 2018). This work provides additional support that young children can represent others’ emotions as jointly caused by their internal mental states and external world states. However, a critical question about the relevance of surprise for learning remains open: Do young children use others’ expressions of surprise as signals of prediction error and use them to guide their own inferences about the world? To answer this question, we ask whether infants (12- to 17-month-olds) can use others’ expressions of surprise to form some expectations about an unobserved outcome. When an agent looks surprised after seeing the outcome of an event (but the outcome is not yet visible to the infant), it provides important information about the nature of the hidden outcome: it is unexpected. Thus, infants would therefore expect an unexpected outcome and look longer when the outcome is expected than unexpected. By contrast, if the agent looks unsurprised, this suggests that the hidden outcome is consistent with what she expected, leading to the usual pattern of looking time (i.e., longer when the outcome is unexpected than when it is expected; Aslin, 2007). Thus, our primary interest here is an interaction effect between the agent’s expression (Surprise vs. No Surprise) and the probability of the outcome (Probable vs. Improbable) in infants’ looking time. 5 We test these predictions using events adapted from a well-established prior study (Xu & Garcia, 2008). This study has shown that 8-month-olds expect randomly drawn samples to be representative of the population; when an agent randomly draws ping-pong balls from a box that contains mostly white and just a few red ones, infants look longer when the sample is mostly red (an improbable outcome) than when it is mostly white (a probable outcome). In our study, we add emotion information to the paradigm; before the agent reveals the outcome to the infant, she first looks at it herself and makes either a surprised or an unsurprised emotional response to it. Then she reveals the outcome to the infant. We predict that while previewing the agent’s unsurprised expression would result in a similar pattern of looking time as the original study, observing a surprised emotional expression would reverse the pattern