/deepcover

DeepCover: Uncover Bugs in Deep Learning

Primary LanguagePython

DeepCover

Uncover Bugs in Deep Learning

DeepCover is a coverage-based bug finder for deep learning applications. It is able to provide metrics for evaluating the robustness of a Deep Neural Network (DNN). The basic idea of DeepCover is easy. For any neuron in the DNN, if its state changes, there must be a cause for it: DeepCover aims to COVER these changes and causes. More details can be found in the paper Testing Deep Neural Networks.

To run the tool, the following software or Python package is required:

theano, numpy, matplotlib, IBM CPLEX