Explaining Predictions of Deep Neural Classifier via Activation Analysis