Scalable Post-Hoc Uncertainty for Deep Networks

S. Rodríguez Santana

Optimal decision-making under uncertainty requires deep neural networks to output robust confidence bounds. Because retraining finely tuned models degrades their carefully optimized accuracy and wastes computational resources, post-hoc Bayesian inference has become an interesting proposal where maintaining performance is crucial. We review this statistical evolution from the Linearized Laplace Approximation (LLA) to accelerated versions (ELLA) and variational approaches. We then introduce the a Gaussian Process-based formulation whose posterior mean is rigidly anchored to the model's exact original forecasts. This restricts stochastic optimization entirely to predictive variances. Bridging statistical rigor and operational efficiency, the method scales independently of dataset size, allowing for new achievements in terms of the calibration of the predictions.

Palabras clave: Approximate inference, Gaussian Processes, Neural Networks

Machine Learning
2 de septiembre de 2026  12:40
Aula 22

D. J. Benito Gutiérrez, J. L. Aguilar Castro, R. E. Lillo Rodriguez

W. Feng, D. Wozabal, C. Ruiz

D. Rodriguez, C. Lalanne, J. Loubes