All times are CET, i,e. local Brussels time. 
The exact schedule is still to be confirmed.

13:00 - 13:05 Electronic Arrive and Welcome
13:05 - 13:45 Thomas Burns: Neurons, Spaces, and Dancing Koalas  
13:50 - 14:30 Anna Juusela, Ulla Koivukoski, Ellimari Kortman & Catharina Vogt:
Access to service by a conversational AI Chatbot for survivors of domestic abuse
14:30 - 14:50 Coffee break and a zoom Group photo
14:50 - 15:30 Patrick Allo: The Virtues and Vices of Formalization - (The Impossible) Formalization of Fairness as a Case Study
15:35 - 16:15 Benedikt Löwe: Conceptual Modelling

The Virtues and Vices of Formalization: (The Impossible) Formalization of Fairness as a Case Study
by Patrick Allo
Within the formal sciences, formalisation and formal rigour are more than just a prerequisite to bring things within the realm of computation or (formal) proof. They are also perceived and presented as valuable in their own right. Aiming at a formally rigorous expression of concepts, claims, or ideas is a way to keep yourself in check. It helps one avoid fallacies and imposes a higher standard of what it means to really understand something.

The social sciences and humanitiesi that are currently most invested in the debate on contemporary data practices (algorithms, Big Data, AI) have a much less positive view of formalisation (Selbst  et al. 2019, Green & Viljoen 2020). While they agree that formalisation is a prerequisite to make things computable and afford correct calculations, they are more likely to associate formalisation with a false sense of understanding. Once presented as a way to avoid human biases, algorithms are now commonly seen as tools that are more likely perpetuate and reinforce existing biases (Mittelstadt et al. 2016): exactly the opposite of seeing formalisation as a means to avoid human (cognitive) biases (Dutilh Novaes 2012).

Against the background of this debate, the virtues of formalisation and the critical potential of formal rigour can be hard to articulate. Formalisation, according to some leading narratives, only serves a narrow technical need and the critical potential of the formal sciences is reduced to its ability to call out bad mathematics (Chiodo & Bursill-Hall 2018).

Critical data studies scholars associate formalisation, much like quantification and the use of mathematical and statistical methods, with positivist epistemologies (Kitchin 2014, Rieder & Simon 2017). One consequence of this association is that the limits of formalisation are more or less equated with the limits of representational conceptions of data. Highly simplified, formalisation is reductive because it presupposes an equally reductive conception of data (see Leonelli & Beaulieu 2021 for the more nuanced version).

My goal in this talk is to explore whether this connection between formalisation and representational conceptions of data plays a similar and equally prominent role in the debate on the (impossibility of the) formalisation of fairness (see e.g., the FAT/ML and FAccT workshop series). One reason to believe that the critique of positivism might not be as central is that the fromalisation of fairness debate is at least in part about the conceptualisation of theoretical and ethical concepts. If this suspicion is correct, it is plausible to ask whether this debate has produced an alternative account of the vices of formalisation. And if we ask this question, we may just as well also ask the opposite question. What do the virtues of formalisation look like when seen from the perspective of this debate? What room is there, when we focus on the formalisation of theoretical concepts, to value formalisation and formal rigour without thereby advancing a positivist conception of the sciences?

***

Neurons, Spaces, and Dancing Koalas   
by Thomas Burns
If we don’t know how the ship sails, how can we hope to steer it away from the rocks? Most recent advances in AI (both in interpretations and applications) rely heavily on vector spaces. Vector spaces are wonderfully versatile mathematical objects, and modern computers make them convenient and efficient to use in the building and interpretation of AI systems. But are they the best and only choice? I think not. In this talk, I will trace some of the historical roots of the modern deep learning formalization doctrines, dismantling them where appropriate and showing they engage in a deep forgetting. I will conclude with an example of my own work, passionately based on curvature in cube complexes.

***

Conceptual Modelling
by Benedikt Löwe

Mathematical modelling is a well-established technique used in the technical and natural sciences in which a real-world phenomenon is modelled by a mathematical representation to make predictions. This general technique has a natural analogue in the humanities and social sciences; its formal representations are often not entirely mathematical or quantitative, but rather conceptual. The name 'conceptual modelling' has been proposed for this analogue. In this talk, we shall explore methodological features of mathematical and conceptual modelling as well as important differences between them.

***

Access to service by a conversational AI Chatbot for survivors of domestic abuse

by Anna Juusela, Ulla Koivukoski, Ellimari Kortman and Catharina Vogt

Victim-survivors of domestic violence typically face many obstacles when seeking and finding help. Fear, shame and fatigue are just a few of the many obstacles that prevent them from coming forward. To lower the threshold for accessing services, the EU-funded IMPRODOVA project is developing the conversational, multilingual AI chatbot AINO. The chatbot is designed to provide immediate advice and risk assessment to victim-survivors and guide them to available service options. The performance of the AI chatbot will be continuously improved through machine learning using input from victim-survivors. The path and current status of development will be presented and discussed. 

For more information, see also: https://www.improve-horizon.eu/