Understanding deep learning via backtracking and deconvolution

Convolutional neural networks are widely adopted for solving problems in image classification. In this work, we aim to gain a better understanding of deep learning through exploring the miss-classified cases in facial and emotion recognitions. Particularly, we propose the backtracking algorithm in order to track down the activated pixels among the last layer of feature maps. We then are able to visualize the facial features that lead to the miss-classifications, by applying the feature tracking algorithm. A comparative analysis of the activated pixels reveals that for the facial recognition, the activations of the common pixels are decisive for the result of classification; for the emotion recognition, the activations of the unique pixels indeed determine the result of classification.