Solving Raven's Progressive Matrices Using RNN Reasoning Network

Although enormous progress has been made by Deep Neural Networks (DNN) in basic perception tasks, they are long criticized for lack of reasoning quality and interpretability. Raven's Progressive Matrices (RPMs) are standard tests for assessing human Intelligence Quotient, also acting as a tool to evaluate Artificial Intelligence. Existing methods of pure DNN combining perception and reasoning are difficult to confirm DNN's ability of logical reasoning. Hybrid DNN methods reasoning by algorithms, not DNN. Existing methods use end-to-end training, while perception and reasoning are detached modules functionally and neurologically. Here we propose a method of separating visual perception by a perfect convolutional neural network perception module abstracting RPMs to panel-level interpretable encoding and inferring it with our Recurrent Neural Network reasoning model. With a trained perfect perception module, our model attaches great performance on the RPM dataset Impartial-RAVEN of 96.72% and RAVEN of 90.77%. Our method tests the abstract reasoning quality of DNN s and also provides a platform for further research in neuroscience.