Neuro-symbolic learning of answer set programs from raw data
File(s)
Author(s)
Cunnington, Daniel
Type
Thesis
Abstract
Artificial Intelligence (AI) is becoming increasingly integrated in society. This raises serious questions of the robustness of the technology. Neuro-Symbolic AI is well-suited to help answer these questions. Many existing neuro-symbolic systems focus on training a neural network given manually engineered background knowledge. However, this can be laborious, and the complete knowledge may not be available. The question is, can we leverage recent advancements in the robustness and expressivity of symbolic machine learning to learn knowledge from raw data? To integrate both neural and symbolic learners, there are various challenges to address: (1) Can exact symbolic knowledge be learned from noisy neural network predictions? (2) Can symbolic knowledge be learned in an end-to-end fashion, since symbolic learners are not differentiable? (3) Can a neuro-symbolic learning framework scale to complex tasks? In this thesis, we propose three novel approaches, capable of learning highly expressive knowledge in the language of Answer Set Programming, defined in terms of primitive concepts extracted from raw data. First, we investigate the robustness of two state-of-the-art symbolic learners when learning complex knowledge, given noisy concept predictions from pre-trained neural networks. Empirical results demonstrate that symbolic knowledge can be learned accurately, even when large percentages of training data are subject to distributional shifts, which causes the networks to predict incorrectly with high confidence. Second, we develop an end-to-end architecture that iteratively trains both neural and symbolic components. We demonstrate that our approach outperforms a variety of baselines, achieving state-of-the-art results. Finally, we explore scalable solutions for neuro-symbolic learning, and develop an architecture that leverages the implicit knowledge embedded within large vision-language foundation models, to extract primitive concepts from raw data. Our evaluation demonstrates the scalability and data efficiency gained with respect to state-of-the-art neuro-symbolic AI methods, in cases where the symbolic knowledge is both given, and learned.
Version
Open Access
Date Issued
2024-10-13
Date Awarded
01/07/2025
Advisor
Russo, Alessandra
Lobo, Jorge
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)