Neural-symbolic learning of interpretable high-level knowledge from raw data

Within the last two decades, Deep Learning has been demonstrated to achieve high accuracy in computational tasks that involve large quantities of data and for which manual feature extraction would be difficult to handle, bringing transformative impact in domains such as facial, image, speech recognition, and natural language processing. But they are still far from showing human-like intelligence. Specially, they are data intensive and lack of transferability, generalisability and interpretability. The recent field of neuro-symbolic AI has recently emerged in response to these shortcomings. However,  most of the recently proposed neural-symbolic methods tend to combine the ability of deep neural networks to learn in noisy, high dimensional environments with the power of symbolic AI mainly to reason over discrete inputs in a logical manner [5].  They therefore require symbolic (or logic-based) knowledge to be manually engineered, limiting their applicability to real-world problems. On the other hand, advancements in logic-based machine learning (also known as inductive logic programmings ) have recently seen the emergence of systems that are scalable and robust to noise and capable of learning from noisy data high-level knowledge, represented in a logic-based form, which is transparent and interpretable (e.g. common-sense knowledge). One of the state-of-the-art logic-based machine learning systems is FastLAS [1] and its recent extended algorithm for learning non-observable concepts [2]. But, a pure hybrid integration of deep neural network with logic-based machine learning remains still an open question. If addressed successfully such hybrid integration has the potential to overcome the pure Deep learning limitations and improve the neural-symbolic AI field by allowing high-level knowledge to be automatically learned from labelled unstructured data.

The goal of this project is to explore a novel end-to-end neural-symbolic learning approach that combines deep learning feature extraction from (multi-modal) unstructured data with logic-based machine learning of high-level knowledge. The project will build on recent results in the SPIKE group, led by Prof Russo, on (i) scalable logic-based machine learning [1,2], (ii) feed-forward neural symbolic learning [3] and (iii) non-monotonic reasoning in continuous vector space [4]. The project will apply the proposed hybrid architecture to a range of tasks, including video question answering, complex event recognition from videos and text, and multi-hop Q&A on knowledge graphs.

[1] Mark Law, Alessandra Russo, Elisa Bertino, Krysia Broda, and Jorge Lobo. FastLAS: Scalable inductive logic programming incorporating domain-specific optimisation criteria. AAAI2020, 34(03):2877–2885.
[2] Mark Law, Alessandra Russo, Krysia Broda, and Elisa Bertino. Scalable Non-observational Predicate Learning in ASP In The Thirtieth International Joint Conference on Artificial Intelligence. IJCAI 2021.
[3] Dan Cunnington, Alessandra Russo, Mark Law, Jorge Lobo, Lance Kaplan. NSL: Hybrid interpretable learning from noisy raw data, 2021.
[4] Yaniv Aspis, Krysia Broda, Alessandra Russo, Jorge Lobo, Stable and Supported Semantics in Continuous Vector Spaces. In Knowledge Representation 2020,
[5] Md Kamruzzaman Sarker, Lu Zhou, Aaron Eberhart, and Pascal Hitzler. Neuro-symbolic Artificial Intelligence: Current trends, 2021,

Project ID



Alessandra Russo


Logic, Norms