Le 22 mars de 15h à 16h
Deep Learning models are being increasingly deployed in real-world applications, where they are likely to be confronted with data that is different from the data they were trained or validated on. It is crucial to be able to identify this situation, especially for safety-critical Machine Learning applications.
The dominant approach to achieving this goal is out-of-distribution (OOD) detection, i.e., identifying when an example comes from the same distribution as the training dataset or not. Oodeel is a library developed by ANITI/DEEL’s team that allows the use of state-of-the-art methods from the OOD detection literature for any image classifier in PyTorch and Tensorflow. It focuses on Post-Hoc OOD methods that apply to already trained models and, therefore, do not require further training.
As a result, it implements lightweight, versatile tools that can be used on any models that are possibly already in production. In this talk, we will showcase some use cases of Oodeel to emphasize its simple, customizable API and the range of tools it offers. We will also present perspectives for the extension to OOD for object detection.
Ce wébinaire s’adresse aux ingénieurs IA, data scientists ou autres MLOps utilisateurs. Il montrera à quoi sert la librairie et comment l’intégrer dans vos plateformes.
French 
English