ICAI: The Labs -Trustworthy AI in the Netherlands
On the 15th of December between 12:00 and 13:00, ‘ICAI: The Labs’ will have a meetup on Trustworthy AI in the Netherlands. Two labs each present their work and discuss challenges and developments in this field.
KPN Responsible AI Lab is a collaboration between KPN and JADS. The lab will focus on developing technologies that allow Artificial Intelligence to be used in a responsible way. The lab will be based in Den Bosch.
Cultural AI Lab bridges the gap between cultural heritage institutes, the humanities, and informatics. It is a collaboration between Centrum Wiskunde & Informatica (CWI), the KNAW Humanities Cluster (KNAW HuC), the National Library of the Netherlands (KB), the Rijksmuseum, the Netherlands Institute for Sound and Vision, TNO, Vrije Universiteit Amsterdam and the University of Amsterdam.
12.00 (noon): Opening
12:05 Gianluigi Bardelloni introduces KPN Responsible AI Lab.
12:10 Rastislav Hronský (Jheronimus Academy, KPN Responsible AI Lab) presents “The Untold Stories of Distributional Semantics in Natural Languages.”
12:25 Laura Hollink (CWI) introduces the Cultural AI Lab.
12:30 Savvina Daniil (CWI, Cultural AI Lab) presents “Responsible Book Recommender Systems”
12.45 Discussion of what’s next in Trustworthy AI in the Netherlands.
“Responsible Book Recommender Systems”
Savvina Daniil is a first year PhD student at CWI (Centrum Wiskunde & Informatica), within the Human-Centered Data Analytics group. Her PhD is funded by KB, the National Library of the Netherlands, with the topic being “Responsible Recommenders in the Public Library sector”.
In this talk, she will be discussing her work around Recommender Systems, a class of algorithms widely used by online platforms, like e-commerce and media content hosts. Specifically, she will present her findings on the unexpected bias that might arise when a system like that is applied in a public library setting.
“The Untold Stories of Distributional Semantics in Natural Languages”
The ideas of distributional structure of language and the associated computational methods have gone through a period of massive development and application, and were eventually overshadowed by deep learning models. While such large models achieve impressive results on most tasks requiring advanced understanding of natural language, their current design has several drawbacks. Large language models come at an extremely high computational and financial cost, their information processing is obscure, and their applicability is predominantly bound to a supervised fine-tuning phase. In addition to requiring expensive labelling of training data, supervised learning fundamentally relies on a well-defined, consensual, and constant set of non-overlapping categories, which is a challenging requirement to meet by itself in practice. We present an approach that integrates a selection of theory and methods in segmentation, distributional semantics, and latent variable models to facilitate unsupervised extraction of high-level, interpretable features that meet the standards set by large language models more economically and transparently. At its core, the approach generalises the conventional, lexical level distributional semantics to a hierarchy of concepts made from abstractions of combinations of lower-level concepts.Such a design is best suited for analysing narrow language domains, where the high-level concepts are low in number of classes and high in terms of re-occurrence per class.