Here you can find the code for our submission to the Siemens AI Dependability Assessment. It solves a binary classification task using a deep neural network and provides security guarantees for local robustness using perturbation analysis based on linear relaxation. Our model displays high predictive performance, gives provable safety guarantees, scales well to more complex data sets, and lets domain experts dynamically configure the class-wise cost of misclassification.
- Patrick Deutschmann, patrick.deutschmann@student.tugraz.at
- Lukas Timpl, lukas.timpl@student.tugraz.at
config.yaml
: the Weights & Biases config file for setting the training parametersrequirements.txt
: all used packages required to run our codetrain.py
: the main script to train the modelnn_model.py
functions to create the modelevaluation.py
: evaluation functionsdata_prep.py
: functions to prepare the data setsbounds_test.py
: a sanity-check test casebaselines.ipynb
: notebook with the baseline computationsdata_prep.ipynb
: notebook investigating the data sets