Raven's Progressive Matrices Completion with Latent Gaussian Process Priors

Abstract reasoning ability is fundamental to human intelligence. It enables humans uncover relations among abstract concepts and further deduce implicit rules from the relations. As a well-known visual task, Raven's Progressive Matrices (RPM) are widely used in IQ tests. Although extensive research has been conducted on RPM solvers with machine intelligence, few studies have considered advancing standard answer-selection (classification) problem more challenging answer-painting (generating) problem, which can verify whether model indeed understood rules. In this paper we aim solve latter one by proposing deep latent variable model, multiple Gaussian processes employed as priors of variables separately learn underlying RPMs; thus proposed interpretable terms concept-specific variables. The process also provides an effective way extrapolation for answer painting based learned concept-changing We evaluate RPM-like datasets continuously-changing concepts. Experimental results demonstrate that our requires only training samples paint high-quality answers, generate novel panels, achieve interpretability through