dl
(Gaussian process classifiers and CNN uncertainty). Convolutional neural networks (CNNs)
achieve state-of-the-art performance on image classification tasks, but provide no measure of
confidence in their predictions. Use the Keras MNIST example CNN available at https://github.com/
fchollet/keras/blob/master/examples/mnist_ cnn.py. After training the model, use the outputs from
the top level feature layer before the softmax classifier to train a Gaussian Process classifier on the
MNIST data. Evaluate your classifier’s performance compared to the original CNN with softmax
classifier. Plot the distribution of classification uncertainties for correctly and incorrectly classified
samples.
Extension: Without re-training, use your model to classify the n-MNIST dataset (http: //csc.lsu.edu/
~saikat/n-mnist/), again plotting the classification uncertainties. Investigate how adversarial
perturbations to the MNIST data affect classification accuracy and uncertainty (try the CleverHans
toolbox for generating adversarial perturbations).