Repairing deep neural networks fix patterns and challenges

http://web.cs.iastate.edu/~design/papers/ICSE-20a/bugrepair.pdf

takeaway

Our key findings reveal that DNN bug fix patterns are distinctive compared to traditional bug fix patterns; the most common bug fix patterns are fixing data dimension and neural network connectivity;

We also contribute a benchmark of 667 DNN (bug, repair) instances.

major challenges:

• DNN bug fixes have the potential to introduce adversarial vulnerabilities;
• DNN bug fixes frequently introduce new bugs;
• DNN bug localization, reuse of trained model
• coping with frequent releases