Skip to content

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.