Project information
- Name: Early-Exit Inspired Dynamic Neural Network For OOD Satellite Imaging
- Category: Machine Learning
- Project date: Sep 2022 - Dec 2022
- Project URL: https://github.com/GM-git-dotcom/la_ood_cs640
Project Detail
- Designed early-exit inspired neural network with One-class SVM classifiers distinguishing Out of Distribution sample by extracting features of So2sat satellite dataset at each layer of a VGG16 backbone.
- Proving layer 2 as best for model-specific OOD detection with 90.14% AUROC and 84.52% AUPR with OOD confidence score as threshold for early exit.