Search for: All records

People are remarkably capable of generating their own goals, beginning with child’s play and continuing into adulthood. Despite considerable empirical and computational work on goals and goal-oriented behaviour, models are still far from capturing the richness of everyday human goals. Here we bridge this gap by collecting a dataset of human-generated playful goals (in the form of scorable, single-player games), modelling them as reward-producing programs and generating novel human-like goals through program synthesis. Reward-producing programs capture the rich semantics of goals through symbolic operations that compose, add temporal constraints and allow program execution on behavioural traces to evaluate progress. To build a generative model of goals, we learn a fitness function over the infinite set of possible goal programs and sample novel goals with a quality-diversity algorithm. Human evaluators found that model-generated goals, when sampled from partitions of program space occupied by human examples, were indistinguishable from human-created games. We also discovered that our model’s internal fitness scores predict games that are evaluated as more fun to play and more human-like. 

A detailed model of the outside world is an essential ingredient of human cognition, enabling us to navigate, form goals, exe- cute plans, and avoid danger. Critically, these world models are flexible—they can arbitrarily expand to introduce previously- undetected objects when new information suggests their pres- ence. Although the number of possible undetected objects is theoretically infinite, people rapidly and accurately infer un- seen objects in everyday situations. How? Here we investigate one approach to characterizing this behavior—as nonparamet- ric clustering over low-level cues—and report preliminary re- sults comparing a computational model to human physical in- ferences from real-world video. 

When people solve problems, they may try multiple invalid solutions before finally having an insight about the correct solution. Insight problem-solving is an example of the flexibility of the human mind which remains unmatched by machines. In this paper, we present a novel experimental paradigm for studying insight problem-solving behavior in a physical reasoning domain. Using this paradigm, we seek to quantify precisely what it means to have an insight during physical problem-solving and identify behavioral traces that predict subjective insight ratings collected from human participants. The project provides the first steps towards a computationally informed theory of insight problems solving. 

A fully occluded object cannot be perceived directly, but we can still infer its existence from the effect it has on the motion and behavior of other, visible objects. Here we report the results of a behavioral experiment designed to elicit these sorts of inferences and quantify their re- liability. Our experiment leverages videos of real-world objects interacting under real-world physics (specifically, interrupted pendulum motion). We propose a preliminary model for how the mind might efficiently infer the position and number of occluded objects simply from the effect they have on the visible physics of a scene.