A framework for automatic identification of neural network structural redundancy based on reinforcement learning

The increasing structure of neural networks makes it difficult to deploy on edge devices with limited computing resources. Network pruning has become one of the most successful model compression methods in recent years. Existing works usually compress models by removing unimportant filters based on the importance. However, the importance-based algorithms tend to ignore the parameters that extract edge features with small criterion values. And recent studies have shown that the existing criteria rely on norm and lead to similar model compression structures. Aiming at the problems of ignoring edge features and manually specifying the pruning rate in current importance-based model pruning algorithms, this paper proposes an automatic recognition framework for neural network structure redundancy based on reinforcement learning. First, we perform cluster analysis on the filters of each layer, and map the filters into a multi-dimensional space to generate similar sets with different functions. We then propose a criterion for identifying redundant filters within similar sets. Finally, we use reinforcement learning to automatically optimize the cluster dimension, and then determine the optimal pruning rate for each layer to reduce the performance loss caused by pruning. Extensive experiments on various benchmark network architectures and datasets demonstrate the effectiveness of our proposed framework.