Using Causal Induction in Humans to Learn and Infer Causality from Video

Abstract

For both human and machine learners, it is a challenge to make high-level sense of observations by identifying causes, effects, and their connections. Once these connections are learned, the knowledge can be used to infer causes and effects where visual data might be partially hidden or ambiguous. In this paper, we present a Bayesian grammar model for human-perceived causal relationships that is learnable from video. Two experiments investigate high-level causal induction from low-level visual cues. In the first experiment, we show that a computer can apply known heuristics used for causal induction by humans to learn perceptual causal relationships. In the second experiment, we show that our learned model can represent humans' performance in reasoning about hidden effects in video, even when the computer initially misdetects those effects.


Back to Table of Contents