IP5: Uncertainty meets Explainability: Combining Uncertainty Quantification and Explainable Machine Learning for Crop Monitoring
Lead: Ribana Roscher
PhD student: Mohamed Farag
Central Question: How to enhance the understanding of DNN models, and improve the robustness of sensing by combining Explainability and Uncertainty Quantification?
Motivation
Trustworthiness, reliability, and robustness in Machine Learning (ML) models become pivotal things to avoid consequential failure when deployed in real life.
Tools
Uncertainty estimation and explainable ML can play a significant role in providing intact outcomes. Let’s define both:
-
- Explainability can provide us with interpretations but in a more human-friendly readable form which reduces the un-easiness of dealing with highly technical concepts for those who are from different backgrounds regarding the model’s predictions.
- Uncertainty quantification is dealing with the wide spectrum of outcomes we would get due to uncertainties already inherited at both data and model (Aleatroic(data)- Epistemic(model) uncertainties).
Work Program
WP1: Uncertainty Quantification
WP2: Uncertainty Calibration
WP3 & WP4: Decomposition of uncertainty sources
Themes’ relation
IP5 contributes and is linked heavily to the other themes of the AID4Crops, it will provide estimates for uncertainty in Theme 1 which would lead to better models and data. Furthermore, for tactical decisions, it would provide outcomes accompanied by uncertainty to guide the policymaker.
