Humans maintain probabilistic belief states when predicting object trajectories

Matjaz Jogan and Alan A Stocker
Computational and Systems Neuroscience meeting CoSyNe, Salt Lake City, March 05-07 2015, Poster presentation.

The ability to accurately predict the trajectories of moving objects is crucial for an autonomous system that interacts with a dynamic environment. Humans are thought of having a sense of so-called "intuitive physics" that allows them to be quite efficient in making such predictions. Recent experimental results suggest that these predictions reflect the outcome of a probabilistic inference process based on noisy observations and an accurate physics model of the world (Smith et al., 2013; Battaglia et al., 2013). However, it remains unknown whether humans mentally track and update i) an estimate or ii) a full probabilistic description of the object state (belief state) (Lee et al., 2014). We designed a set of psychophysical experiments to specifically distinguish the two hypotheses. Subjects were first asked to predict the collision location of a moving object with a hidden wall. The trajectory of the object was occluded and subjects were only given the object's initial motion and an acoustic signal at the precise time of collision. Subjects exhibited clear biases in their location estimates that indicated that they were performing probabilistic inference using prior expectations over speed and location. Subjects then repeated the experiment receiving, however, an additional spatial cue about the hidden wall location. By introducing different levels of uncertainty associated with this cue we expected subjects to assign different relative weights in combining the cues if they were maintaining full belief states while tracking. More specifically, by measuring subjects performance for each cue alone we were able to individually predict optimal behavior and verify whether it matched subjects' actual behavior. We found that subjects' behavior was indeed well predicted by a Bayesian belief state model that optimally combined cues across space, time, and object motion. Our results suggest that humans maintain and update full belief states when predicting object trajectories.