Workshop on Human-Interpretable AI Gabriele Ciravegna Mateo Espinoza Zarlenga Pietro Barbiero gabriele.ciravegna@polito.it University of Cambridge Università della Svizzera Italiana Dipartimento di Automatica e Cambridge, UK Lugano, Switzerland Informatica, Politecnico di Torino Torino, Italy Francesco Giannini Zoreh Shams Damien Garreau Scuola Normale Superiore University of Cambridge Julius-Maximilians-Universität Pisa, Italy Cambridge, UK Würzburg Würzburg, Germany Mateja Jamnik Tania Cerquitelli University of Cambridge Dipartimento di Automatica e Cambridge, UK Informatica, Politecnico di Torino Torino, Italy Abstract 1 Introduction This workshop aims to spearhead research on Human-Interpretable Human-interpretable AI models [1] are playing an increasingly Artificial Intelligence (HI-AI) by providing: (i) a general overview important role in Artificial Intelligence (AI). Today, a large part of of the key aspects of HI-AI, in order to equip all researchers with the technologies employed by AI and SIGKDD researchers is based the necessary background and set of definitions; (ii) novel and on Deep Neural Networks (DNNs). Yet, the lack of transparency of interesting ideas coming from both invited talks and top paper DNNs prevents a safe deployment of these models in critical con- contributions; (iii) the chance to engage in dialogue with promi- texts that significantly affect users. Consequently, decision-making nent scientists during poster presentations and coffee breaks. The systems based on deep learning are facing constraints and limita- workshop welcomes contributions covering novel interpretable- tions from regulatory institutions [2], which increasingly demand by-design or post-hoc approaches, as well as theoretical analysis transparency in AI models [3]. Even though standard eXplainable of existing works. Additionally, we accept visionary contributions AI (XAI) emerged to address the need to interpret DNNs, several speculating on the future potential of this field. Finally, we welcome works are arguing that it may not have achieved its goal [4, 5]. contributions from related fields such as Ethical AI, Knowledge- To really explain DNN decision-making process, there is a grow- driven Machine learning, Human-machine Interaction, applications ing consensus that human-interpretable explanations are required. in Medicine and Industry, and analyses from Regulatory experts. Human-Interpretable AI (HI-AI) methods either provide post-hoc explanations by extracting the symbols that have been automati- CCS Concepts cally learnt by the models (e.g., T-CAV [6]), or directly design in- trinsically interpretable architectures (e.g., CBM [7]). Among other • Computing methodologies → Artificial intelligence. qualities, these explanations resemble better the way humans rea- son and explain [8], help to detect model biases [9], are more stable Keywords to perturbations [10], and can create more robust models [11]. Human-Interpretable AI, Interpretability, Explainability, HI-AI, XAI ACM Reference Format: 2 Workshop Topics Gabriele Ciravegna, Mateo Espinoza Zarlenga, Pietro Barbiero, Francesco Gi- Topics of interest include, but are not limited to, the following: annini, Zoreh Shams, Damien Garreau, Mateja Jamnik, and Tania Cerquitelli. • Explainable-by-design models, novel approaches to cre- 2024. Workshop on Human-Interpretable AI. In Proceedings of the 30th ACM ating machine learning and deep learning models that are SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24), intrinsically explainable or interpretable. August 25–29, 2024, Barcelona, Spain. ACM, New York, NY, USA, 2 pages. • Post-hoc methods for Interpretable AI, novel approaches https://doi.org/10.1145/3637528.3671499 on post-hoc interpretable AI. These include but are not lim- ited to approaches working on higher-level features such as Permission to make digital or hard copies of all or part of this work for personal or concepts. classroom use is granted without fee provided that copies are not made or distributed • Theoretical analyses of existing methods, showing what for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. existing interpretable methods can achieve both from an For all other uses, contact the owner/author(s). explanation and a generalization point of view. KDD ’24, August 25–29, 2024, Barcelona, Spain • Knowledge integration & Reasoning methods injecting © 2024 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-0490-1/24/08 domain knowledge and reasoning methods into deep learn- https://doi.org/10.1145/3637528.3671499 ing models to enhance their interpretability and performance. CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings KDD ’24, August 25–29, 2024, Barcelona, Spain Gabriele Ciravegna et al. • AI Ethics papers analysing implications of interpretable AI 8:50 – 9:00 Opening remarks methods, discussing topics such as fairness, accountability, 9:00 – 9:40 Keynote: Andrea Passerini transparency, and bias mitigation in AI systems. 9:40 – 10:00 5 mins lightning talks (3 selected papers) • Human-machine Interaction studies on innovative human- 10:00 – 10:40 Keynote: Abbas Rahimi machine interaction systems, successfully exploiting inter- 10:40 – 11:30 Coffee & Posters pretable AI models in their capability to provide both stan- 11:30 – 12:10 Keynote: Sonali Parbhoo dard and counter-factual explanations. 12:10 – 12:20 Awards and Closing Remarks • Vision papers on XAI discussing the possible evolutions of Table 1: Draft of the program outline. the XAI field or speculating potential interpretable system and applications with their implications. • Applications in Medicine and Healthcare applications of interpretable AI methods in medical diagnosis, treatment In the case of research contributions, we asked paper authors to planning, and healthcare decision-making. make their code and data openly available to ensure reproducibility. • AI in Industry practical applications of interpretable AI The review process has been double-blind. We have used OpenRe- methods in various safety-critical industrial sectors, such as view to ensure the final decisions for each paper are made by the transportation, finance and retail. organisers with no conflict of interest. All accepted papers will be • Legal and Regulatory dissertations discussing and pro- published on the workshop website, which will remain active and viding analysis of the legal challenges associated with inter- accessible after the conference concludes. Additionally, we took pretable AI, including compliance with data protection laws contact with an external editor (CEUR-WS) to create an archival for transparent and accountable AI systems. version of these papers for authors who wish to participate in a subsequent publication. 3 Program 5 Program Commitee This workshop aims to advance the research on HI-AI by offering a diverse program designed to enhance participants’ knowledge, We are very grateful to each of our program committee members and foster collaboration and innovation. The following list contains for their hard reviewing work, namely Romain Giot, Eliana Pastor, the invited speakers who will give keynote talks at the HI-AI work- Roberto Pellungrini, Eleonora Poeta, Gianluigi Lopardo, and Gizem shop, and the expected topics that their talks will cover. All invited Gezici, besides workshop chairs. speakers have already confirmed their presence. References • Abbas Rahimi, Research Staff Member at IBM Research [1] Eleonora Poeta, Gabriele Ciravegna, Eliana Pastor, Tania Cerquitelli, and Elena Europe - Neuro-symbolic AI, Concept Embeddings. Baralis. Concept-based explainable artificial intelligence: A survey. arXiv preprint • Andrea Passerini, Associate Professor at University of arXiv:2312.12936, 2023. [2] Bryce Goodman and Seth Flaxman. European union regulations on algorithmic Trento - Concepts in AI and Interactive Machine Learning. decision-making and a “right to explanation”. AI magazine, 38(3):50–57, 2017. • Sonali Parbhoo, Assistant Professor at Imperial College [3] Johann Laux, Sandra Wachter, and Brent Mittelstadt. Trustworthy artificial intelligence and the european union ai act: On the conflation of trustworthiness London - Concept and causality. and acceptability of risk. Regulation & Governance, 18(1):3–32, 2024. [4] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, Program Outline. Table 1 reports the workshop program. Firstly, and Been Kim. Sanity checks for saliency maps. Advances in neural information we will give an overview of the key aspects of HI-AI to ensure all processing systems, 31, 2018. attendees have a solid understanding of the background concepts [5] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence, and terminology. Secondly, the workshop features three invited 1(5):206–215, 2019. talks from experts in the field, who will share their insights and lat- [6] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing est research findings. These talks will provide valuable perspectives with concept activation vectors (tcav). In International conference on machine and inspire new ideas. Thirdly, we will offer participants the chance learning, pages 2668–2677. PMLR, 2018. to engage in dialogue with prominent scientists during a long coffee [7] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pier- son, Been Kim, and Percy Liang. Concept bottleneck models. In International break with poster presentations, encouraging collaborations and conference on machine learning, pages 5338–5348. PMLR, 2020. knowledge-sharing. Also, the workshop program includes three [8] Sunnie SY Kim, Elizabeth Anne Watkins, Olga Russakovsky, Ruth Fong, and contributed talks from selected contributions. We will recognize the Andrés Monroy-Hernández. " help me help the ai": Understanding how ex- plainability can support human-ai interaction. In Proceedings of the 2023 CHI most interesting contribution with a Best Workshop Paper Award. Conference on Human Factors in Computing Systems, pages 1–17, 2023. We have allocated 40 minutes for each invited talk, allowing for a [9] Rishabh Jain, Gabriele Ciravegna, Pietro Barbiero, Francesco Giannini, Davide Buffelli, and Pietro Lio. Extending logic explained networks to text classification. 30-minute presentation followed by a 10-minute Q&A session. We In Proceedings of the 2022 Conference on Empirical Methods in Natural Language allotted the same time for the poster sessions. Processing, pages 8838–8857. Association for Computational Linguistics, 2022. [10] David Alvarez Melis and Tommi Jaakkola. Towards robust interpretability with self-explaining neural networks. Advances in neural information processing sys- 4 Paper Management tems, 31, 2018. Paper management. We published the Call For Papers (CFP) on [11] Gabriele Ciravegna, Pietro Barbiero, Francesco Giannini, Marco Gori, Pietro Lió, Marco Maggini, and Stefano Melacci. Logic explained networks. Artificial the workshop website1 . The CFP focuses on short papers, which Intelligence, 314:103822, 2023. can be research papers, theoretical analysis papers, or vision papers. 1 https://human-interpretable-ai.github.io/