In this work, we study the power of Saak features as an effort towards\ninterpretable deep learning. Being inspired by the operations of convolutional\nlayers of convolutional neural networks, multi-stage Saak transform was\nproposed. Based on this foundation, we provide an in-depth examination on Saak\nfeatures, which are coefficients of the Saak transform, by analyzing their\nproperties through visualization and demonstrating their applications in image\nclassification. Being similar to CNN features, Saak features at later stages\nhave larger receptive fields, yet they are obtained in a one-pass feedforward\nmanner without backpropagation. The whole feature extraction process is\ntransparent and is of extremely low complexity. The discriminant power of Saak\nfeatures is demonstrated, and their classification performance in three\nwell-known datasets (namely, MNIST, CIFAR-10 and STL-10) is shown by\nexperimental results.\n