Current advances in Artificial Intelligence (AI) and Machine Learning (ML) have achieved unprecedented impact across research communities and industry. Nevertheless, serious concerns around trust, safety, interpretability and accountability in AI were raised by influential thinkers. Many identified the need for well-founded knowledge representation and reasoning to be integrated with ML systems. Neurosymbolic AI has been an active area of research for many years seeking to do just that, bringing together robust learning in neural networks with reasoning and explainability via symbolic representation and description. In this talk I will review the research in neurosymbolic AI and computation, and discuss how it can help shed light into the increasingly prominent role of safety, trust, interpretability and accountability in AI. AI has become the focus of large-scale research endeavours and has changed businesses. This led to an important debate about the impact of AI on education and society. It has been argued that the building of a rich AI system, semantically sound, explainable and ultimately trustworthy, will require a sound reasoning layer in combination with deep learning. Parallels have been drawn between Daniel Kahneman’s research on human reasoning and decision making and so-called AI systems 1 and 2. I will revisit early theoretical results of fundamental relevance to shaping the latest research, such as the proof that recurrent neural networks compute the semantics of logic programming. I will also seek to identify bottlenecks and the most promising technical directions for the sound representation of learning and reasoning in neural networks. I will conclude by discussing the key ingredients for sustainable AI going forward, identifying directions and challenges for the next decade of research in the field.

Artur d'Avila Garcez is Professor of Computer Science at City, University of London. He is a Fellow of the British Computer Society (FBCS), Fellow of the UK’s Higher Education Academy (FHEA), and president of the steering committee of the Neural-Symbolic Learning and Reasoning association (NeSy). He has co-authored two books: Neural-Symbolic Cognitive Reasoning, Springer (2009) and Neural-Symbolic Learning Systems, Springer (2002), and co-founded the NeSy conference series, the longest standing conference series in neurosymbolic AI. Garcez is a world leader in the field of neurosymbolic AI. He designed and implemented arguably the first neurosymbolic system for learning and reasoning. His research in the area has led to more than 200 publications in major journals and conferences in Artificial Intelligence, Logic and Machine Learning, and in the flagship AI and Neural Computation conferences AAAI, NeurIPS, IJCAI, IJCNN, AAMAS and ECAI. Garcez is a founding Editor-in-Chief of the Neurosymbolic AI journal, IOS Press, and holds editorial and senior program committee positions in all the major scientific journals and conferences in Artificial Intelligence and Machine Learning.

The advent of Large Language Models (LLMs) like GPT-4 represents a significant leap forward in the field of Artificial Intelligence, offering unprecedented capabilities in understanding, generating, and interacting with human language. This talk aims to demystify these complex systems, explaining their fundamental architecture, how they are trained on vast datasets, and the underlying technologies that enable their sophisticated processing abilities. We will explore the importance of LLMs, highlighting their role in driving innovation, enhancing productivity, and opening new avenues for human-computer interaction. However, with great power comes great responsibility, and LLMs are not without their limitations and challenges. This presentation will critically examine the inherent weaknesses of LLMs, such as biases in training data, the potential for generating misleading information, and ethical concerns. We will delve into real-world examples to illustrate these failures, offering a balanced perspective on the capabilities and limitations of these models. Finally, I will overview ongoing research in my group aimed at mitigating these shortcomings, including extracting facts from generated text and interconnecting them to the source text.

Disclaimer: 90% of this abstract was generated by GPT-4.

Roberto Navigli is Professor of Natural Language Processing at the Sapienza University of Rome, where he leads the Sapienza NLP Group, and an ACL Fellow. He has received two prestigious ERC grants in AI on multilingual word sense disambiguation (2011-2016) and multilingual language- and syntax-independent open-text unified representations (2017-2022), highlighted among the 15 projects through which the ERC transformed science. In 2015 he received the META prize for groundbreaking work in overcoming language barriers with BabelNet, a project also highlighted in The Guardian and Time magazine, and winner of the Artificial Intelligence Journal prominent paper award 2017 (and a subsequent AIJ prominent paper award in 2023 on the NASARI sense embeddings). He is the co-founder of Babelscape, a successful deep-tech company which enables NLU in dozens of languages. He served as Associate Editor of the Artificial Intelligence Journal (2013-2020) and Program Chair of ACL-IJCNLP 2021. He will serve as General Chair of ACL 2025.