An Interpretable Neuro-symbolic Model for Raven's Progressive Matrices Reasoning

Raven's Progressive Matrices (RPM) have been widely used as standard intelligence tests for human participants. Humans solve RPM problems in a hierarchical manner, perceiving conceptual features at different levels and inferring the latent rules governing the matrix using cognitive maps. Although the latest AI algorithms can surpass human performance, little effort has been made to build a model that solves RPM problems in a human-like hierarchical manner. We built a human-like hierarchical neuro-symbolic model to solve RPM problems. The proposed model consists of a semantic-VAE (sVAE) perceptual module and a cognitive map reasoning back-end (CMRB). The supervised sVAE extracts the hierarchical visual features of RPMs by perceiving the structural organization of RPMs through a convolutional neural network and disentangles objects into semantically understandable features. Based on these semantic features, the CMRB predicts the semantic features of objects in the missing field using cognitive maps generated by supervised learning or manually designed. The answer image was generated by sVAE using the semantic features predicted by CMRB. The proposed model achieved state-of-the-art performance on three benchmarks datasets-RAVEN, I-RAVEN, and RAVEN-fair-generalizes well to RPMs containing objects with untrained feature dimensions, mimics human cognitive processes when solving RPM problems, achieves interpretability of their hierarchical processes, and can also be applied to some real-world situations that require abstract visual reasoning.