RADMI: Latent Information Aggregation as a Proxy for Model Uncertainty

2026-05-02

AI Summary

This work addresses the challenge of efficiently estimating epistemic uncertainty in dense prediction tasks with deep learning by proposing a single forward-pass uncertainty estimation method based on inter-layer mutual information within the decoder. By computing normalized mutual information between consecutive decoder layers in an encoder-decoder architecture and linearly aggregating the resulting information flow to quantify prediction reliability, the approach generates spatially precise uncertainty maps without requiring architectural modifications, ensembles, or multiple inference passes. Notably, it introduces inter-decoder-layer mutual information as a novel proxy for epistemic uncertainty. Evaluated on a seismic facies segmentation benchmark, the method substantially outperforms existing single-pass approaches, improving Pearson and Spearman correlation coefficients by 5.5% and 10.7%, respectively, and achieves uncertainty calibration performance comparable to that of deep ensemble methods.

Abstract

Epistemic uncertainty estimation is essential for identifying regions where deep learning system outputs may be unreliable. However, existing approaches require computationally expensive ensemble methods or multiple stochastic forward passes, limiting their scalability to dense prediction tasks like segmentation. We propose Resolution-Aggregated Decoder Mutual Information (RADMI), a single-pass method that estimates prediction uncertainty by measuring mutual information (MI) between consecutive decoder layers in segmentation networks. We observe that elevated inter-layer MI correlates with prediction uncertainty, as the network must integrate conflicting contextual information at ambiguous regions such as class boundaries. Evaluating on a seismic facies segmentation benchmark, RADMI achieves the highest correlation with deep ensemble uncertainty among all single-pass methods, outperforming the next-best baselines by 5.5% in Pearson and 10.7% in Spearman correlation coefficients. Compared to baselines that either lack spatial precision or demand significant computational overhead, RADMI yields sharp, boundary-localized uncertainty maps without architectural modifications. Our results suggest that linear aggregation of normalized information flow provides a principled and efficient proxy for prediction uncertainty in encoder-decoder architectures.

Problem

• epistemic uncertainty
• dense prediction
• segmentation
• model uncertainty
• computational efficiency

Innovation

• epistemic uncertainty
• mutual information
• single-pass uncertainty estimation
• semantic segmentation
• decoder aggregation