Tutorial: Last-Layer VI for Heat2D Failure Classification¶
Notebook: LastLayerVI_Heat2D_Classification_Tutorial.ipynb
Scientific problem¶
Classify 2D heat-source maps into safe and failure regimes based on whether the peak temperature exceeds a fixed threshold.
Input and output¶
- input: 2D source field
- output: binary class probabilities for
safe/failure
UQ method¶
- deterministic convolutional feature extractor
LastLayerVariationalInference(..., task="classification")predict_vi_uq(...)returns class probabilities and probability variance
Why this notebook exists¶
This notebook is the scalable VI example for larger deterministic backbones. Only the final linear head is Bayesian, which keeps training and sampling practical.