preview Final project of the Neural Networks course based on the work Interpreting CNNs via Decision Trees [1].

Convolutional neural networks are nowadays widely used for different tasks in many fields, becomes thus important to understand what knowledge a CNN learns. In order to do so we have modified a VGG-16 network by adding a special mask layer at the end of the convolution, and trained it for an image classification task. Finally we have built a decision tree that could help us explain which object parts contributed the most to the final prediction and quantify these contributions

GitLab repo

[1]: Q. Zhang, Y. Yang, H. Ma and Y. Wu, “Interpreting CNNs via Decision Trees,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019 pp. 6254-6263. doi: 10.1109/CVPR.2019.00642