Published May 12, 2025
Library
graduated
Neural DE
DockerContainer
PythonPackage
Maintainer:IRT-SystemX
Description
NeuralDE is a library made to improve the robustness of your models at test time. It proposes a set of methods that will allow you to remove identified corruptions in your data before you send it to your model. This library addresses issues such as meteorological corruptions, distribution shifts etc.
Owners:Air LiquideIRT-SystemX
Keywords:neuralde
CONTEXT
Real input data to machine learning models can be deliberately perturbed or come from a context different from the one used during training, leading the model to make incorrect predictions.
VALUE PROPOSITION
NeuralDE offers several classes of transformations methods:
- For pre-processing images for Adversarial purification (annihilating adversarial noise from inputs),
- Image Restoration (Remove noise, meteorological weather corruptions, digital corruptions),
- Domain Shift Inversion (shifting data back to the original context).
Transformations will be used to pre-process input images before being used in inference by the AI component
WHEN TO USE IT
It will be used by Data Engineers or ML-Algorithm Engineers during the Deploy and maintain ML-based System activity.
RESOURCES
