-
Train a FFN (fedforward neural network), for example fully connected network, CNN, and ResNet, with MNIST dataset.
-
Select one image from each class from the training set.
-
Save the images and their logits from the first nn.
-
Use data augmentation and generative models to train a second FFN network.
In file "FFN.ipynb", the basic FFN is fully connected network. The best ACC score is 0.72.
In file "ResNet.ipynb", the basic FFN is ResNet18. The best ACC score is 0.81, the best AUC score is 0.95.