MASTERCLASS MINT 2026 | Sébastien Gerchinovitz | March 26, 2026

Item: MASTERCLASS MINT 2026 | Sébastien Gerchinovitz | March 26, 2026
Rating: 4
Author: PostSingularity Institute

by Sébastien Gerchinovitz

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning

🎙 Sébastien Gerchinovitz 👥 182 📅 April 7, 2026 ⏱ 69 min 👁 77 📄 lecture 🧭 2026-08-15

Available in: English (current) Français

Keywords

conformal predictionsplit conformalexchangeabilitynonconformity scoreprediction sets

Summary

This masterclass, given by Sébastien Gerchinovitz, provides an introduction to conformal prediction, a statistical framework for quantifying uncertainty in supervised learning. The talk begins by motivating the need for uncertainty quantification, citing examples such as plant recognition, magnetic declination prediction, and object detection for self-driving cars. It then introduces the concept of a nonconformity score, which measures the error between a prediction and the true label. The main focus is on split conformal prediction, a simple three-step protocol: train a model on training data, compute nonconformity scores on a separate calibration set, and then construct prediction sets for new inputs based on a quantile of these scores. The speaker presents a theorem guaranteeing that the true label will be contained in the prediction set with probability at least 1-alpha, under the assumption that the data are exchangeable. The proof is sketched using symmetry arguments. The talk concludes with a brief discussion of limitations and extensions, and mentions an application to computer vision.

162 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a clear and valuable introduction to conformal prediction, a topic of increasing importance in machine learning. The speaker effectively motivates the need for uncertainty quantification, especially in high-stakes applications. The argumentation is solid: the method is presented step-by-step, and the theoretical guarantee is stated and proved in a way that is accessible to a non-specialist. The use of examples helps to illustrate the concepts. The speaker also acknowledges limitations and extensions, which adds to the credibility of the presentation.

Scientific Rigor, Source Quality, Title Accuracy

The talk is scientifically rigorous, with a clear statement of assumptions and a formal theorem. The speaker attributes the method to Vovk et al. (2002) and Vovk (2005), and mentions that similar arguments appeared in the 1940s. The title accurately reflects the content. The talk does not rely on external sources beyond the cited references, but the description provides a link to the event page on Indico, which may contain additional resources. The presentation is well-structured and the mathematical arguments are sound.

179 words

Title / Content Match

The title accurately reflects the content: an introduction to conformal prediction.

Quality & Reliability

8/10

The talk is a clear, rigorous introduction to conformal prediction, with a formal theorem and proof. The speaker is a researcher at IRT Saint-Exupéry, and the content is well-structured. However, the video is a lecture, not a peer-reviewed publication, and the proof is sketched rather than fully detailed.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Event page on Indico — The talk was part of the MINT Masterclass series; the page may contain slides or additional materials.

Concurring Sources

Conformal prediction (Wikipedia) — General overview consistent with the talk's content.

Contribution & Novelties

The talk provides a clear and accessible introduction to conformal prediction, a method that is often presented in a more technical manner. It emphasizes the simplicity of the approach and its practical applicability, contrasting it with more complex deep learning theory. The proof is presented in an intuitive way, highlighting the role of exchangeability. The talk also mentions limitations and extensions, giving a balanced view.

Pour aller plus loin :

Conformal prediction (Wikipedia) — Overview of the field.
Vovk et al., 2005, Algorithmic Learning in a Random World — Foundational book.
Angelopoulos & Bates, 2021, A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification — Accessible tutorial.

108 words

Radar Profile

The radar profile shows high scores in information quantity, quality, and reliability, with a slightly lower technical level, indicating a well-balanced and accessible presentation.

Reliability 8/10