Mitigating fooling with competitive overcomplete output layer neural networks

Denoising Autoencoders for Overgeneralization in Neural Networks

“…Fooling of each target digit was attempted by performing stochastic gradient descent on the parameters of the FGN to minimize the cross-entropy between the output of the network to be fooled and the specific desired target output class. Fooling of each digit was attempted for 20 trials with different random inputs to the FGN, each trial consisting of up to 10000 parameter updates, as described in [7].…”

“…All the models were trained for a fixed 100 epochs. Fooling was attempted at two different thresholds, 90% and 99%, in contrast to the previous work that used only the 99% one [7]. Comparing the models at different thresholds indeed gives more information about their robustness and may amplify their differences, thus improving the comparison.…”

“…The three models that were tested on fooling, a plain CNN, COOL [7] and the dAE model described in this paper were next compared in the context of open set recognition.…”

“…The limit of open set recognition in which a single training class is observed during training, that is the problem of 1-class recognition, was next explored, comparing the model presented in this paper with COOL [7] and 1-Class SVM [14].…”

“…A separate 1-class recognition problem was created from the MNIST and Fashion-MNIST datasets for each target class. For each problem the models were trained using only samples from the selected , for three models, the dAE model described in this paper, 1-Class SVM [14] and COOL [7].…”