FFNSL: Feed-Forward Neural-Symbolic Learner

Abstract Logic-based machine learning aims to learn general, interpretable knowledge in a data-efficient manner. However, labelled data must be specified structured logical form. To address this limitation, we propose neural-symbolic framework, called Feed-Forward Neural-Symbolic Learner (FFNSL) , that integrates logic-based system capable of from noisy examples, with neural networks, order unstructured data. We demonstrate the generality FFNSL on four classification problems, where different pre-trained network models and systems are integrated sequences images. evaluate robustness our framework by using images subject distributional shifts, for which networks may predict incorrectly high confidence. analyse impact these shifts have accuracy learned run-time performance, comparing tree-based pure approaches. Our experimental results show outperforms baselines more accurate fewer examples.