{"id":7,"url":"https://pm.philipcastiglione.com/papers/7.json","title":"Abductive Knowledge Induction From Raw Data","read":false,"authors":"Wang-Zhou Dai, Stephen Muggleton","year":2021,"auto_summary":"The paper \"Abductive Knowledge Induction From Raw Data\" by Wang-Zhou Dai and Stephen Muggleton, presented at the IJCAI-21 conference, introduces a novel approach called Abductive Meta-Interpretive Learning (MetaAbd). This method addresses the challenge of integrating sub-symbolic perception with symbolic reasoning in neuro-symbolic AI, specifically focusing on the issue of where symbolic knowledge originates from in learning systems.\n\nMetaAbd combines abduction and induction to jointly learn neural networks and induce logic theories from raw data. This approach allows for the learning of recursive first-order logic theories with predicate invention, which can be reused as background knowledge in future tasks. The system leverages abductive reasoning to infer the most probable explanations for observed data, thereby reducing the search space for potential hypotheses and improving computational efficiency.\n\nThe paper highlights the limitations of existing neuro-symbolic models, which often assume a strong pre-existing symbolic knowledge base and focus only on learning perception models. In contrast, MetaAbd does not require such assumptions and can learn both the perception and reasoning components simultaneously.\n\nExperimental results demonstrate that MetaAbd outperforms state-of-the-art end-to-end deep learning models and neuro-symbolic methods in tasks involving cumulative sum and product calculations from images of handwritten digits, as well as sorting tasks. MetaAbd achieves higher predictive accuracy, better data efficiency, and learns interpretable models that can generalize to longer input sequences.\n\nThe paper also discusses the implementation details of MetaAbd, including the use of a meta-interpreter for logic program induction and the integration of probabilistic abduction to guide the learning process. The authors emphasize the potential of MetaAbd to bridge the gap between machine learning and logical reasoning, offering a promising direction for future research in neuro-symbolic AI.","notes":{"id":6,"name":"notes","body":null,"record_type":"Paper","record_id":7,"created_at":"2024-12-10T03:51:40.518Z","updated_at":"2024-12-10T03:51:40.518Z"},"created_at":"2024-12-10T03:51:33.702Z","updated_at":"2024-12-10T03:51:46.343Z"}