Learned Uncertainty-Aware (LUNA) Bases for Bayesian Regression using Multi-Headed Auxiliary Networks
The original LUNA paper can be found here.
FinalReport.ipynb
- summary of the original LUNA paper, evaluation of experimental design, and remarks on our process of replicating the authors' results
feed_forward.py
- Contains the base class for a neural network, adapted from in-class codenlm.py
- Contains the base class for the NLM model which defines a neural linear modelluna.py
- Contains the base class for the LUNA model; inherits fromnlm.py
bayes_helpers.py
- Helper functions for Bayesian analysis within the LUNA model; contains functions for sampling from the prior or posterior, calculating the prior/posterior predictive, plotting predictive intervals, calculatingutils.py
- helper functions for neural network components of the LUNA model; contains functions for generating training data (with a gap) and running a toy neural net/plotting resultsconfig.py
- Contains standardized configuration parameters for NLM and LUNA models
LUNABaseDemo.ipynb
- demonstration of LUNA model on a toy datasetPriorPredictives_Demo.ipynb
- demonstration of how regularization affects the prior and posterior predictive of an NLMCategoricalFailure.ipynb
- demonstration of a LUNA failure mode: uncertainty estimation for 2-D classification
- Directory figs_final - Contains plots of true data, predictions, and predictive uncertainty for NLM and LUNA training examples.