Tipo di tesi
Tesi di laurea magistrale
Titolo
Development of post-hoc weight perturbation ensemble models for out-of-distribution image detection
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Parole chiave
- cross-entropy calibration
- deep ensembles
- ensemble diversity
- out-of-distribution detection
- post-hoc methods
- uncertainty estimation
- weight perturbation
Data inizio appello
15/04/2026
Consultabilità
Non consultabile
Data di rilascio
15/04/2096
Riassunto (Inglese)
Out-of-Distribution (OOD) detection is essential for deploying deep neural networks in safety-critical settings. Most post-hoc OOD methods analyze a single model output (e.g., softmax confidence, energy, or feature-space distances), without probing prediction stability under model perturbations. This thesis introduces a post-hoc method based on stochastic weight perturbation ensembles. Starting from one pre-trained network, we generate an ensemble at inference time by injecting layer-wise Gaussian noise into model parameters, with per-layer scaling proportional to weight standard deviation and controlled by a global coefficient λ. OOD scoring uses predictive entropy of the ensemble mean prediction; in-distribution samples remain stable across perturbations, while OOD samples induce larger variability. The main contribution is an ID-only analytical calibration criterion for λ. Using the ensemble loss decomposition, we show that in-distribution ensemble cross-entropy exhibits a characteristic cradle-shaped trend and define λ* as its zero-crossing point. This yields automatic calibration without OOD validation data or manual tuning. Experiments on OpenOOD across CIFAR-10, CIFAR-100, ImageNet-200, and ImageNet-1K, with ResNet-18, ResNet-50, ViT-B/16, and Swin-T, show competitive AUROC and FPR@95 versus strong baselines, while preserving the simplicity and practicality of a post-hoc approach.
