The Development of Joint Belief-Desire Inferences

The Development of Joint Belief-Desire Inferences Hilary L. Richardson 1 (hlrich@mit.edu), Chris L. Baker 1 (clbaker@mit.edu), Joshua B. Tenenbaum 1 (jbt@mit.edu), and Rebecca R. Saxe 1,2 (saxe@mit.edu) Department of Brain and Cognitive Sciences 1 and McGovern Institute for Brain Research 2 Massachusetts Institute of Technology Cambridge, MA 02139 USA Abstract work studying the development of representational ToM has used versions of the False-Belief task to ask children to predict an agent’s behavior, given the agent’s previously established beliefs and desires. In contrast, there has been less work studying how children solve the inverse problem: inferring beliefs and preferences from an observed behavior. Given the girl’s extensive search for a fruit in the kitchen, how do we simultaneously infer her beliefs and preferences? Both kinds of judgments—predicting actions given beliefs and desires, and inferring beliefs and desires given actions—tap similar ToM reasoning abilities. This paper aims to test whether the development of the ability to make ToM inferences parallels the transition to understanding false beliefs, and to provide evidence for a formal account of the knowledge supporting both ToM abilities. The ability to solve this inverse problem is analogous to solving one equation with two unknowns; our natural ability to consider context, weigh in with priors, and make rational inferences enables us to come up with a good guess on questions we would otherwise not be able to answer. Studying this ability in the social domain illustrates the power of ToM to go “beyond the data” and infer multiple implicit mental states from just one observed action. Prior work by Baker et al. (2011) presented adult participants with an inverse mental-state-inference task and showed that adult mental state inferences are well-explained within a rational probabilistic inference framework. Here, we use an analogous paradigm to measure spontaneous mental state inferences made by children 3-6 years of age and assess which observed behaviors prompt mental state inference. By doing so, we are measuring children’s expectation that all parts of an action should have a sufficient explanation in terms of mental states. If this inferential ability develops in parallel with the ability to predict behavior given a mental state, we would expect a similar shift in performance between the ages of three to five on inverse problems that require mental state inference. On the other hand, it is possible that the ability to infer mental states from sparse information develops later in life. This process not only requires the ability to take the perspective of another and maintain multiple representations of the world, but it also requires that the viewer spontaneously seeks to understand observed actions in terms of underlying beliefs and desires. In our experiment, children watched a short 3D animation of a hungry bunny navigating a world to find and eat one of three different fruits. The bunny can take one of three paths: (1) pass the nearest, visible fruit to check around a wall to choose the fruit there (2) take a direct path to the nearest, Human beings infer complex mental states given very little information — a facial expression, a sarcastic tone, or even a simple behavior. Previous work suggests that adults make joint belief and desire inferences based on an actor’s path, and that these inferences are well-explained by a Bayesian framework (Baker, Saxe, & Tenenbaum, 2011). We investigate the development of this ability by assessing mental state inferences made by children ages 3-6 after watching a short movie. Our results suggest that young children spontaneously make inductive inferences about desires or preferences, and that the ability to infer belief from behavior develops between ages 3-6, and possibly throughout later childhood. We formulate three computational models that capture the developmental shift between non- representational and representational theory of mind, and show that these models capture qualitative patterns in the children’s data. Keywords: theory of mind, false-belief task, Bayesian inference, cognitive development Introduction As we move about the world, our actions are the observable manifestation of unobservable intentions: we act to fulfill our hopes and desires in accordance with our beliefs. Adults understand this intuitively. When a girl exclaims “I’m starving—I’m craving a piece of fruit!” and begins to search extensively for an apple in the kitchen even though a pear is in plain sight, adults can infer that the girl wants to eat a fruit, that she has a preference for apples over pears, and that she has a reasonable degree of belief that there is an apple in the kitchen. Our explanation of the girl’s action in terms of inferred beliefs and desires relies on a Theory of Mind (ToM): we understand that agents have a working representation of the world that may or may not reflect reality, that this representation is influenced by perceptual access and priors, and that this representation is the basis for subsequent behavior (Gopnik & Wellman, 1992). If the girl believes that apples are in the fruit basket, we confidently predict she will look for one there, even if we know that the apples are actually in the cupboard. This ability is assessed by the famous “False-Belief task 1 ” (Wimmer & Perner, 1983), on which children typically transition from failure to success between the ages of three and five (Wellman, Cross, & Watson, 2001). Most prior “Explicit” vs. “implicit”: Surprisingly, infants succeed in looking-time paradigms tapping analogous notions of perceptual access and false belief representation (Onishi & Baillargeon,