Genetic Algorithm-Based Structure Reduction for Convolutional Neural Network

Due to the heavy computational burden, deep models of embedded and mobile systems inevitably require network reduction with minimum performance degradation. Pruning method is mainly used to reduce the model by removing some filters only within the layer without changing the structure. Some methods for structural reduction of models are far from optimization. We propose a structure reduction method using a genetic algorithm to optimize the removal of reducible layers. Knowledge distillation is carried out to recover the resultant network. We evaluate our method for ResNet on two image classification datasets, CIFAR-10 and CIFAR-100. Experiments show that our method performs a significant improvement over other state-of-the-art methods.