Mini-workshop on the philosophy of machine learning
21 December 2023 - 21 December 2023
Room C2.09, Sector C, East Campus USI-SUPSI
This mini-workshop on the philosophy of machine learning sees the participation of a group of selected scholars who will discuss topics such as the role of justification in machine learning algorithms and a novel role for explainable artificial intelligence as a "mediator". More information and details in the programme here below.

The programme of the workshop

10.00 - 11.00:  Talk by Juan Duran (TU Delft)

Justification, reliabilism, and machine learning

In this talk, I aim to explore the significance of justification in machine learning (ML). To begin, I'll briefly touch upon two promising epistemologies for ML—transparency and computational reliabilism (CR). However, my focus will be on defending the latter, requiring a more in-depth discussion. I'll dedicate some time to elucidate how CR operates, and which assumptions are built-in. Next, I plan to illustrate how CR works in the context of Forensic ML. This emerging field sparks debates about justification due to the inherent challenges in achieving both explanation and understanding, which are crucial for judicial and forensic purposes. Lastly, I'll address two objections against CR: i) the concern that, under CR, statistically insignificant yet serious errors can compromise the reliability of AI algorithms; and ii) the argument that CR, being a reliabilist epistemology, demands a high frequency of success, ultimately posing an issue of high predictive accuracy. I'll present arguments to counter these objections, advocating for computational reliabilism as a promising epistemology for ML.


11.00 - 12.00: Joint talk by Emanuele Ratti (Bristol) and Alberto Termine (IDSIA USI-SUPSI).

On the mediating role of XAI in scientific research

In recent years, philosophers have deserved increasing attention to the eXplainable AI (XAI) research programme. Surprisingly, their analyses focused more on the typologies and format of explanations delivered by XAI models, while little has been said about the potential epistemic roles of XAI in scientific research. In this talk, we will provide a novel framework to understand XAI in scientific research as a class of models that, rather than explaining, mediates. We call this framework ‘XAI-as-mediator’, and we will build on the literature on ‘models as mediators’ to pinpoint its characteristics. In particular, we argue that XAI models mediate between opaque models generated by ML algorithms and the domain/theoretical knowledge of the field to which models are applied.
In support to our thesis, we will analyse two examples of how XAI models can play the role of mediators. The first example focuses on the use of a specific family of XAI tools, namely feature selection methods, as support tools for post-hoc evaluation of ML models by domain experts. The second example, on the other hand, examines the use of XAI, and specifically counterfactual explanation methods, as tools to support hypothesis formulation during exploratory research. The talk is based on a joint work with Alessandro Facchini  (IDSIA USI-SUPSI)


14.00 - 14.45: Talk by Andrea Ferrario (ETHZ)

The quest for justification in AI - where do we stand?

Establishing well-grounded beliefs about artificial intelligence (AI) and its capabilities is crucial, considering its widespread applications. However, existing research lacks a unified approach to the justification of these beliefs. I briefly review current literature, identifying challenges, and suggesting a few insights to advance the field.


14.45 - 15.45: Talk by Chiara Manganini (UniMI)

On the Ontology of Machine Learning Systems and its consequences for the taxonomy of Miscomputation

When compared to "traditional" computational artefacts, ML systems show a crucial difference in the role played by their function, discovered through the training process, rather than fixed from the beginning. This has deep implications for the notions of correctness and, consequently, of miscomputation in ML. By adapting the Ontology of the Levels of Abstraction, a revised framework is proposed to accommodate the essential features of ML systems. The result is a complex ontology composed of three artefacts: the Training Sample, the Training Engine, and the Machine Learning Model. This new ontological framework is then used to develop systematic insights into the types of ML errors, their relationship with non-ML miscomputations, and with fairness and explainability. The talk is based on joint works with Alberto Termine (IDSIA USI-SUPSI), and Giuseppe Primiero (UniMI).

The speakers

Juan Duran is Assistant Professor at the Faculty of Technology, Policy and Management, TU Delft. His research focuses on the philosophy of science and ethics of computer-based science and engineering (computer simulations, AI, and Big Data).
He is the 2019 recipient of the Herbert A. Simon Award for outstanding research in computing and philosophy.  

Andrea Ferrario is the Scientific Director of the Mobiliar Lab for Analytics at ETH and PostDoc at ETH Zurich. His research interests lie at the intersection of philosophy and technology, with a focus on the philosophy of AI and health interventions.

Chiara Manganini is PhD student at the University of Milan, Department of Philosophy. She is part of the Logic, Uncertainty, Computation and Information Group, where she studies the logical and philosophical aspects of the problem of bias in machine learning. 

Emanuele Ratti is Lecturer (i.e. tenure-track) in the Department of Philosophy at the University of Bristol. His areas of specialisation are the History and Philosophy of Science and Technology (molecular biology, genomics, and AI), and Ethics of Science and Technology (including virtue ethics).

Alberto Termine is Assistant Researcher at IDSIA USI-SUPSI. He obtained a PhD at the Logic, Uncertainty, Computation and Information Lab, Department of Philosophy, University of Milan. His current research spans from causal and counterfactual methods in Explainable Artificial Intelligence to the metaphysics and epistemology of Machine Learning.

st.wwwsupsi@supsi.ch