We are glad to announce the 1st european Knowledge-Guided Machine learning workshop, to be held in Turin, Italy in ECML-PKDD 2023.
All the information about the conference can be found at the official website.
Even though machine learning (ML) and deep learning (DL) algorithms have achieved amazing results in
many commercial and business applications, data-driven models have so far met with limited success in
many scientific domains. Limitations of data-driven models arise due to their intrinsic black-box nature,
their large data requirements, their inability to produce physically consistent results, and the lack of
generalizability to out-of-sample scenarios. More generally, the popularity and success of ML-based systems
has put the spotlight on issues such as explainability, bias, fairness, and sustainability. There is thus a
growing interest in developing Knowledge Guided Machine Learning (KGML) approaches that can leverage
decades (sometimes centuries) of accumulated scientific knowledge and combine them with ML techniques
to reap the benefits of both approaches. Addressing these issues naturally leads to systems in which
representation of prior and domain knowledge in various forms, from physical and simulation models to
symbolic and logical representations, plays a central role and must be integrated into ML models and
pipelines. On the other hand, ML also offers solutions for long-standing challenges in the field of Knowledge
Representation (KR), for instance related to efficient, neurally-guided, noise-tolerant and ampliative
inference, knowledge acquisition, efficient reasoning, and limitations of existing physical and symbolic
models. The synergy between ML and KR has the potential to lead to new advancements in fundamental AI
challenges including, but not limited to, learning symbolic generalizations from raw (multi-modal) data,
using knowledge to facilitate data-efficient learning, speeding up inference, supporting interpretability of
learned outcomes and integration of symbolic planning and reinforcement learning.
The workshop aims to provide researchers and practitioners with a dedicated forum for discussing new ideas and research results at the intersection of machine learning and knowledge representation, both from a theoretical and application standpoint.
The workshop will be held on September 22nd, 2023, from 9:00 a.m. to 13:00 a.m. (Italian time) at Room 4i of the Main Campus of the Politecnico di Torino ( map here).Papers denoted with are Extended abstracts from recently published works. The associated full papers are linked below.
|9:00 a.m. - 9:05 a.m.||WELCOME, Introduction by the chairs|
|9:05 a.m. - 9:20 a.m.||
|9:20 a.m. - 9:35 a.m.||
SEEDOT: Tool for Enhancing Sentiment Lexicon with Machine Learning
Vittorio Haardt, Lorenzo Malandri, Fabio Mercorio, Luca Porcelli
|9:35 a.m. - 9:50 a.m.||
Lorentz-invariant augmentation for high-energy physics
Simone Monaco, Sebastiano Barresi, Daniele Apiletti
|9:50 a.m. - 10:05 a.m.||
Unsupervised Ontology- and Taxonomy Construction through Hyperbolic Relational Domains and Ranges
Filip Cornell, Yifei Jin, Jussi Karlgren, Sarunas Girdzijauskas
|10:05 a.m. - 10:20 a.m.||
Discovering SpatioTemporal Warning Contexts from Non-Emergency Call Reports
Luca Cagliero, Andrea Avignone, Silvia Chiusano
|10:20 a.m. - 10:35 a.m.||
A Filter-based Neural ODE Approach for Modelling Natural Systems with Prior Knowledge Constraints
Cecília Coelho, M. Fernanda P. Costa, Luís L. Ferrás
|10:35 a.m. - 10:45 a.m.||
Prior Knowledge Meets Neural ODEs: A Two-stage Training Method for Improved Explainability
Cecília Coelho, M. Fernanda P. Costa, Luís L. Ferrás Full paper
|11:00 a.m. - 11:30 a.m.||Coffee Break|
|11:30 a.m. - 12:00 a.m.||
Keynote - Neurosymbolic AI: Some pain, a lot of gain!
Tarek R. BesoldBio
When introducing neurosymbolic integration to researchers, one is often met with the question: “But why bother? Can’t I just machine learn it?”. Surprisingly often – particularly, when considering the current popularity of ML (and particularly DL) in academic research and the popular press – the answer is “No, you can’t.”. In this talk, we will have a look at the main motivation(s) for why integrating between symbolic and subsymbolic computation is often desirable, if not unavoidable. We will have a look at several use cases (taken from a variety of domains ranging from biomedical research to multisensory perception and context-adaptive robot control) which as of today appear to be almost impossible to solve without making recourse to both ML- and logic-based approaches. We will identify some of the reasons why a combination between data and knowledge is required, and will consider some of the most common ways of how to combine symbolic and subsymbolic aspects in solving the corresponding challenges.
|12:00 a.m. - 12:15 a.m.||
Domain Knowledge Alleviates Adversarial Attacks in Multi-Label Classifiers
Stefano Melacci, Gabriele Ciravegna, Angelo Sotgiu et al. Full paper
|12:15 a.m. - 12:30 a.m.||
Not All Neuro-Symbolic Concepts Are Created Equal: Analysis and Mitigation of Reasoning Shortcuts
Emanuele Marconato, Stefano Teso, Antonio Vergari, Andrea Passerini
|12:30 a.m. - 12:45 a.m.||
Symbolic Regression via Control Variable Genetic Programming
Nan Jiang, Yexiang Xue Full paper
|12:45 a.m. - 13:00 a.m.||Panel and Closing remarks|
Tarek R. Besold is a Senior Research Scientist at Sony AI in Barcelona leading the Gastronomy Flagship Project, as well as an affiliated researcher with the Philosophy & Ethics group at Eindhoven University of Technology. His work covers topics at the intersection between AI, cognitive science, and real-world applications. He also serves as a start-up advisor and consulting AI policy expert, and as chairman of the German DIN/DKE Standards Working Committee on AI NA 043-01-42 GA.
Authors are invited to submit articles combining aspects of domain knowledge representation
and ML, including the use of knowledge-aware methods for solving ML challenges (e.g. knowledge-guided
or explainable learning), the use of ML methods for solving-knowledge representation challenges (e.g.
efficient inference, knowledge base completion), the integration of learning and reasoning at the modeling
or solving side, and the application of combined knowledge-driven ML approaches to solve real-world
problems and accelerate scientific discovery.
We welcome papers on a wide range of topics, including but not limited to:
All deadlines are 11:59PM UTC-12:00 ("anywhere on Earth").
Submissions should present original results not currently published or under review elsewhere. This workshop accepts:
All submissions must be written in English and formatted according to the Springer LNCS style [Paper template]. The submissions must be in PDF format.
The Workshops and Tutorials will be included in a joint Post-Workshop proceeding published by Springer Communications in Computer and Information Science, in 1-2 volumes, organised by focused scope and possibly indexed by WOS.
Authors will have the faculty to opt-in or opt-out.
You can contact the organizers at email@example.com