FF-NSL: Feed-Forward Neural-Symbolic Learner

Abstract

Inductive Logic Programming (ILP) aims to learn generalised, interpretable hypotheses in a data-efficient manner. However, current ILP systems require training examples to be specified in a structured logical form. To address this problem, this paper proposes a neural-symbolic learning framework, called Feed-Forward Neural-Symbolic Learner (FF-NSL), that integrates state-of-the-art ILP systems, based on the Answer Set semantics, with Neural Networks (NNs), in order to learn interpretable hypotheses from labelled unstructured data. To demonstrate the generality and robustness of FF-NSL, we use two datasets subject to distributional shifts, for which pre-trained NNs may give incorrect predictions with high confidence. Experimental results show that FF-NSL outperforms tree-based and neural-based approaches by learning more accurate and interpretable hypotheses with fewer examples.

Inductive Learning of Complex Knowledge from Raw Data

“…However, as these systems are logic-based, they can only learn from structured data. Even recent differentiable methods [29,32,33] are only applied to structured data, and use pretrained neural networks when applied to raw data [7,14,15]. To address this limitation, M eta Abd [8] is the first neuro-symbolic learning approach that uses abduction and induction to jointly train a neural network and induce logic programs from raw data.…”

“…The ILASP system is free to use for research, 5 FastLAS 6 and the FashionMNIST dataset 7 are both open-source with an MIT license, the MNIST dataset is licensed with Creative Commons Attribution-Share Alike 3.0, 8 and the CNN models used from DeepProbLog are open-source and licensed with Apache 2.0. 9…”

“…For example, many existing approaches improve the training of a neural network using a symbolic knowledge base that is manually engineered [2,25,26,30,38]. In contrast, systems that learn symbolic knowledge are generally only applied to structured data [5, 27-29, 32, 33], and use pretrained neural networks when applied to raw data [7,14,15]. To address this limitation, M eta Abd jointly trains a neural network whilst learning symbolic knowledge [8], using a meta-interpreter learner that only learns first-order definite logic programs [6].…”