A novel arbitrary-position solver for Raven's Progressive Matrix Problems

To advance the capacity of abstract reasoning in artificial systems, our research focuses on tackling Raven's Progressive Matrices (RPM) using neural networks, given the high correlation between solving RPM puzzles and human intelligence. Unlike existing solvers for RPM, which are limited to predicting the missing image at the bottom-right position and often overlook relations across rows and columns, we introduce a novel approach that considers the target position, thus enabling the prediction of the missing image at any position within the matrix. Specifically, we introduce a multi-scale encoder to extract multi-resolution representations encompassing both local and global information. Furthermore, we employ Element-wise Attention mechanism to facilitate logical reasoning. Experimental results on the RAVEN and RAVEN-FAIR datasets demonstrate that our proposed method surpasses current state-of-the-art methods and exhibits robust generalization performance across different problem sets.