Commonsense Psychology in Human Infants and Machines

Abstract

Artificial intelligence (AI) promises to take the flawed intelligence of humans out of machines. Why, then, might we want to put the inchoate intelligence of human infants into machines? While infants seem to intuit others’ underlying intentions merely by observing their actions, AI systems, in contrast, fall short in such commonsense psychology. Here we put infant and machine intelligence into direct dialogue through their performance on the Baby Intuitions Benchmark (BIB), a comprehensive suite of tasks probing commonsense psychology. Following a preregistered design and analysis plan, we collected 288 individual responses of 11-month-old infants to BIB’s six tasks and tested three state-of-the-art learning-driven neural-network models from two different model classes. Infants’ performance revealed their comprehensive understanding of agents as rational and goal-directed, but the models failed to capture infants’ knowledge. These striking differences between human and artificial intelligence are critical to address to build machine common sense.

From infancy, we model people as minds, acting in a physical world

“…Two classic behavioral findings from the infant cognition literature show that infants are sensitive to the motives and physical constraints that guide people's actions. First, infants appear to represent other agents' actions with respect to particular goal objects (Feiman et al, 2015;Luo & Baillargeon, 2005;Shimizu & Johnson, 2004;Stojnic et al, 2022;Woodward, 1998). After seeing an agent choose to reach for or move towards one of two objects repeatedly, and the objects switch locations, infants look longer when the agent moves on the same path towards a new object, than when the agent moves on a new path towards the same object.…”

“…After seeing an agent choose to reach for or move towards one of two objects repeatedly, and the objects switch locations, infants look longer when the agent moves on the same path towards a new object, than when the agent moves on a new path towards the same object. Second, infants appear to represent these goal-directed behaviors as constrained by the physical world (Csibra, 2008;Gergely et al, 1995;Phillips & Wellman, 2005;Stojnic et al, 2022). In these experiments, infants first see an agent navigate over or around an obstacle towards a single goal object.…”

“…The causal models presented in these papers use probabilistic generative models of action planning and physical simulation to judge the plausibility of expected and surprising outcomes, given familiarization videos. The feedforward predictive models tested on the same datasets performed reasonably well when tested on scenarios similar to those from the training set, but could not handle scenarios outside of the training distribution (Shu et al, 2021;Stojnic et al, 2022;Zhi-Xuan et al, 2022).…”

“…This allows us to compare responses from babies and models given the very same input. In one paper, Stojnic et al (2022) compare plausibility judgments from a suite of computational models to behavioral responses to the same stimuli in babies. They found, using a novel stimulus set inspired by developmental psychology (the Baby Intuitions Benchmark or BIB; Gandhi et al, 2021), that infants look longer when an agent fails to follow the same goal object when it changes location (but not when the agent's preferred object was occluded at test , or when a new agent is presented with the same options), and when an agent 2 takes an inefficient path towards a single goal object (but not when the agent initially behaved inefficiently).…”

“…The computational models they tested, in contrast, showed a different pattern from these successes and failures. This suite of models did not include probabilistic generative models that express the main hypothesis of this paper; however, Stojnic et al (2022) provide a paradigmatic example of computational cognitive science and developmental psychology working together to test and develop hypotheses about the nature of the infant mind using a shared task and stimulus set.…”