Modelling a Haptic Categorization Task using Bayesian Inference

Abstract

We rely heavily on our sense of touch to complete a myriad of tasks each day, yet past research focuses heavily on the visual and auditory systems, rarely concentrating on the tactile system. In the current study, we investigate human performance on a haptic categorization task and ask: what strategy do humans use to sense, interpret, and categorize objects using their sense of touch? During the experiment, participants complete 810 trials on which they receive a 3D printed object and categorize it as belonging to Category A or B. We sample the objects from a set of 25 objects, each of which differs in number of sides and dot spacing on one face. We define Categories A and B using overlapping Gaussian distributions, where Category A objects generally have fewer sides and smaller dot spacing, while Category B objects generally have more sides and larger dot spacing. Participants begin with no knowledge of the categories and learn them using feedback provided on each trial. We compared human performance to a Feature-Focused Bayesian Observer that weights the sides and dots feature information based on their reliability. It combines information from one or both features to inform a final percept and categorize each object. Our results support the hypothesis that humans employ a feature-focused categorization strategy on this task, during which they learn the categories and consider one or both of an object’s features based on their reliability. As participants complete more trials, they appear to maintain or switch to more optimal categorization strategies. Video analysis of hand movements during the experiment strongly supports these findings.