Self-Supervised Learning for Symbolic Manipulation with Recurrent Neural Networks and Graph Representation Learning

Abstract

This paper presents a novel approach to symbolic manipulation learning using self-supervised learning techniques. We leverage Recurrent Neural Networks (RNNs) to learn manipulation rules directly from data, combined with Graph Representation Learning (GRL) to capture the structural complexities inherent in symbolic domains. The core idea is to train an RNN to predict the outcome of applying a sequence of symbolic operations on a graph representation of the input, eliminating the need for explicit, human-annotated labels. This approach enables RNNs to develop sophisticated reasoning capabilities without relying on traditional supervised learning paradigms. We argue that the combination of RNNs' sequential processing power and GRL's ability to encode structural information provides a strong foundation for learning symbolic manipulation, potentially unlocking more robust and adaptable systems for automated reasoning. The effectiveness of this strategy is demonstrated through the framework's design and theoretical justification, paving the way for future research and development in this area.