arXiv:2506.18283v2 Announce Type: replace 
Abstract: Neural networks make accurate predictions but often fail to provide reliable uncertainty estimates, especially under covariate distribution shifts between training and testing. To address this problem, we propose a Bayesian framework for uncertainty estimation that explicitly accounts for covariate shifts. While conventional approaches rely on fixed priors, the key idea of our method is an adaptive prior, conditioned on both training and new covariates. This prior naturally increases uncertainty for inputs that lie far from the training distribution in regions where predictive performance is likely to degrade. To efficiently approximate the resulting posterior predictive distribution, we employ amortized variational inference. Finally, we construct synthetic environments by drawing small bootstrap samples from the training data, simulating a range of plausible covariate shift using only the original dataset. We evaluate our method on both synthetic and real-world data. It yields substantially improved uncertainty estimates under distribution shifts.

تم اقتراح إطار بايزي جديد لتقدير عدم اليقين في الشبكات العصبية، حيث يتناول التحدي المتمثل في إجراء تنبؤات موثوقة في ظل تغيرات توزيع المتغيرات. تقدم هذه الطريقة أولوية تكيفية تزيد من عدم اليقين للمدخلات التي تنحرف عن توزيع التدريب، مما يحسن الأداء التنبؤي.

Se ha propuesto un nuevo marco bayesiano para la estimación de incertidumbre en redes neuronales, abordando el desafío de realizar predicciones fiables bajo cambios en la distribución de covariables. Este método introduce un prior adaptativo que aumenta la incertidumbre para entradas que se desvían de la distribución de entrenamiento, mejorando así el rendimiento predictivo.

Un nouveau cadre bayésien pour l'estimation de l'incertitude dans les réseaux de neurones a été proposé, abordant le défi des prédictions fiables en cas de changements de distribution des covariables. Cette méthode introduit un prior adaptatif qui augmente l'incertitude pour les entrées s'écartant de la distribution d'entraînement, améliorant ainsi la performance prédictive.

A new Bayesian framework for uncertainty estimation in neural networks has been proposed, addressing the challenge of reliable predictions under covariate distribution shifts. This method introduces an adaptive prior that increases uncertainty for inputs deviating from the training distribution, enhancing predictive performance.

Quantifying Uncertainty in the Presence of Distribution Shifts

Was this article worth reading? Share it

LucidQuery AI

Hypertune

Portfolio Backtest

The Visualizer

Meteoria

Https

Ready to build your own newsroom?