1. Raven solver: From perception to reasoning.
- Author
-
Wei, Qinglai, Chen, Diancheng, Yuan, Beiming, and Ye, Peijun
- Subjects
- *
ARTIFICIAL intelligence , *COMPUTER vision , *VISUAL learning , *DEEP learning , *MODEL-based reasoning , *ECLECTICISM - Abstract
A higher level of machine intelligence is expected after the prosperity of deep learning in computer vision. The Raven's progressive matrices (RPM), for example, raises the requirements of learning from perception to reasoning. Previous solutions for RPM tend to build monolithic end-to-end models, whose excellent performances are of no comfort given that their interpretability and generalizability are far from satisfactory. With hindsight, we formalize RPM as a stratified problem, and propose an eclectic solver. To recapitulate, we explore the possibility of deep contrastive learning for interpretable feature extraction, deep regression with rule representation, and information pooling for reasoning and generalization. Our model achieves 94.63% and 92.90% in RAVEN and I-RAVEN datasets, and demonstrates interpretability in visual attribute learning. More than that, we also discover one potential defect of the original RAVEN dataset. • Develop a modular RAVEN solver in a bottom-up and top-down manner, instead of a monolithic end-to-end model. • Consider the model interpretability in terms of visual attribute extraction and rule discovery in abstract reasoning. • Detect one potential defect of the original RAVEN dataset. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF