=Paper=
{{Paper
|id=Vol-3841/preface
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-3841/Workshop.pdf
|volume=Vol-3841
|authors=Gabriele Ciravegna,Mateo Espiniza Zarlenga,Pietro Barbiero,Francesco Giannini,Zohreh Shams,Damien Garreau,Mateja Jamnik,Tania Cerquitelli
}}
==None==
https://ceur-ws.org/Vol-3841/Workshop.pdf
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/