Fraunhofer uncertainty metrics for classification tasks
Dockerized AI4EU Acumos component for uncertainty estimation for classification networks
Main Characteristic
This is a dockerized component for the AI4EU Acumos platform implementing uncertainty estimation metrics for classification tasks.
Research areas
Verifiable AI
Technical Categories
Machine learning
Last updated
25.05.2021 - 21:29
Detailed Description
Hardware architecture: X64
Additional information: The input to the metric computation module is a prediction from multiple forward passes of Monte Carlo Dropout or models in an ensemble. The prediction is expected as a single data point, so the shape is N x C where N is the number of forward passes, and C is the number of classes.
Documents
Trustworthy AI
This is a meta tool for AI. It supports determining the uncertainty of an AI generated prediction and thereby aims at increasing the "Technical robustness and safety" and the "Transparency" of the analysed AI.
GDPR Requirements
The algorithm is not containing or using GDPR relevant data in itself.