Cascaded reasoning neural networks for artificial general intelligence

Abstract

This paper introduced a new approach to computational reasoning using an LSTM that has four inputs representing four different sources including one from a previous network and another from a knowledge base, called CRN. The network is able to extract the relevant information from the inputs and compress them to a concise, symbolic output that can become an input for the next network. This network thereby makes successive inferences in a cascaded manner, achieving a chain of reasoning with the end result as the answer to the main question, similar to how humans perform step by step reasoning to solve complex problems. CRN is also able to command the computer to compute a formula or provide more information. The study demonstrated the efficacy of CRN by solving a modified version of the bAbI QA database. After the augmentation of the dataset with labels for intermediate reasoning, CRN achieved the score of 20/20 on all tasks in the bAbI, being first to do so.

Keywords: Artificia, Intelligence, Language, Reasoning, Neural, Networks, LSTM