CNN各層による異常検出

抄録

In this paper, we analyze how different layers look upon the anomalies in test data by extracting outputs of each hidden layer in a Convolutional Neural Network (CNN). To this end, we make different kinds of anomalous data to put into a CNN. Among them, some are like the training data, while others are completely different. By feeding those anomalous data into the MNIST-based pre-trained CNN, we can get the corresponding outputs at each hidden layer, and then we can find out which features in the hidden layers of the CNN will be treated as anomalies.