Stratified Rule-Aware Network for Abstract Visual Reasoning

Abstract reasoning refers to the ability analyze information, discover rules at an intangible level, and solve problems in innovative ways. Raven's Progressive Matrices (RPM) test is typically used examine capability of abstract reasoning. The subject asked identify correct choice from answer set fill missing panel bottom right RPM (e.g., a 3×3 matrix), following underlying inside matrix. Recent studies, taking advantage Convolutional Neural Networks (CNNs), have achieved encouraging progress accomplish test. However, they partly ignore necessary inductive biases solver, such as order sensitivity within each row/column incremental rule induction. To address this problem, paper we propose Stratified Rule-Aware Network (SRAN) generate embeddings for two input sequences. Our SRAN learns multiple granularity different levels, incrementally integrates stratified embedding flows through gated fusion module. With help embeddings, similarity metric applied guarantee that can not only be trained using tuplet loss but also infer best efficiently. We further point out severe defects existing popular RAVEN dataset test, which prevent fair evaluation ability. fix defects, generation algorithm called Attribute Bisection Tree (ABT), forming improved named Impartial-RAVEN (I-RAVEN short). Extensive experiments are conducted on both PGM I-RAVEN datasets, showing our outperforms state-of-the-art models by considerable margin.