Convolutional neural network simplification with progressive retraining

Kernel pruning methods have been proposed to speed up, simplify, and improve explanation of convolutional neural network (CNN) models. However, the effectiveness a simplified model is often below original one. In this letter, we present new based on objective subjective relevance criteria for kernel elimination in layer-by-layer fashion. During process, CNN retrained only when current layer entirely simplified, by adjusting weights from next first one preserving subsequent layers not involved process. We call strategy \emph{progressive retraining}, differently that usually retrain entire after each simplification action -- e.g., or few kernels. Our criterion exploits ability humans recognizing visual patterns improves designer's understanding The combination suitable progressive retraining shows our can increase with considerable simplification. also demonstrate provide better results than two popular ones another state-of-the-art using four challenging image datasets.