Deep Learning Confidence Calibration Methods Developed for Noisy Data and Privacy

This thesis explores methods for improving the confidence calibration of deep learning systems, particularly in high-stakes applications where reliable uncertainty quantification is crucial. It addresses challenges such as label noise and domain shifts by proposing a framework that reconstructs noise-free confidence estimates using an estimated noise model. The research also extends to unsupervised domain adaptation, enabling calibration without labeled target data by estimating target-domain accuracy from source performance and domain discrepancies. Furthermore, it introduces a locally differentially private conformal prediction framework to ensure privacy while maintaining reliable uncertainty quantification. AI

IMPACT Enhances the reliability and safety of deep learning models in critical applications by improving uncertainty quantification.

RANK_REASON The cluster contains an academic paper detailing new research methods for deep learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• artificial neural network
• Conformal Prediction
• deep learning

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →