Uncertainty quantification for deep learning techniques

Rouaa Hoblos

Lundi 4 Mars 2024 à 16h30
SUPMICROTECH-ENSMM – Amphi Jules Haag

Most of the real physical system and everyday situations include uncertainty. This is the case for medical diagnosis, weather forecasting, evolution of the stock market and so on. In the literature two types of uncertainty are distinguished: aleatoric uncertainty denotes the one that is inherent to the data, e.g., noise in measurements or natural variability of the inputs, and epistemic uncertainty related to the model and due to lack of knowledge. Measuring the uncertainty is important, so as to support the user in the action to take.
For example, when an anomaly is detected, with weak confidence level, another source of information should be added (image, human intervention, etc.) before planning intervention actions. More generally, quantification of the prediction uncertainty allows to trust or not predictions.
In fact, incorrect overconfident predictions can be harmful and lead to erroneous decision.

I will present main methods of uncertainty quantification for deep learning, with detailed application related to breast cancer using neural networks.