This is our tentative program. More details coming soon.
Tutorial: On the Evolution of Interpretable Methods
- 09h00-09h10: Welcome
- 09h10-10h30: [Introduction] – [Early Interpretable Methods]
- 10h30-11h00: Coffee Break
- 11h00-12h40: [From Classical to Mechanistic Interpretability] – [From Mechanistic to Compositional Interpretability]
- 12h40-12h55: Discussion and Concluding Remarks.
- 12h00-14h00: Lunch Break
Workshop
- 14h00-15h00 (Keynote Talk): “Improving the Intelligibility of Explanations: Some Methods and Risks” by Marie-Jeanne Lesot
- 15h00-15h15 (Paper presentation): Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations
- 15h15-15h30 (Paper presentation): COCOLogic-V2: Identifying Logical Inconsistencies via Truly Hard-Negatives
- 15h30-15h45 (Paper presentation): Evaluating Meta-Feature Fidelity: Detecting Information Leakage in Interpretable Latent Spaces, by Charlotte Job
- 15h45-16h00 (Paper presentation): How smoothing the affinity matrix affects neighborhood preservation in t-SNE
- 16h00-16h30: Coffee Break
- 16h30-16h45 (Paper presentation): Hallucination Neurons and Where to Find Them: An Investigation into the existence of Hallucination Neurons
- 16h45-17h00 (Paper presentation): When Models Disagree: Contrastive Explainability for Clinical Mortality Prediction
- 17h15-18h00: Poster Session
Keynote Talk: Improving the Intelligibility of Explanations: Some Methods and Risks
Explanations generated by XAI methods can take many different forms, e.g. feature importance scores, data influence scores, counterfactual data points, natural language texts or logical formulas to name a few. Depending on the addressed explainee, some explanations may actually be more or less understandable, although intelligibility is an obviously crucial requirement explanations need to satisfy. The talk will focus on two families of XAI methods expected to be easily understood: the first one relies on the expression of the explanation as natural language, in particular in the case of LLM self-explanations. The second one relies on the integration of tools from the fuzzy set theory, that has been designed to allow for interpretability and legibility. The talk will discuss some risks that may arise in each case, underlining the need for careful studies of the characteristics and quality of the generated explanations, despite the difficulty of these tasks.
