An Iterative Deep Neural Network Pruning Method Based on Filter Redundancy Assessment

Storage and inference of deep neural network models are resource-intensive, limiting their deployment on edge devices. Structured pruning methods can reduce the resource requirements for model storage and inference by selectively removing low-contributing filters. Current methods do not focus on the functional relations between filters and cannot remove filters that are scored as high-contributing but have redundant feature extraction capabilities. To address this limitation, we propose a structured pruning method, Important filter Decreasing Iteratively (IDI), which removes low-contributing filters and functionally redundant filters in two pruning stages, respectively. The first stage of the IDI removes low contribution filters using a filter contribution evaluation criterion. In the second stage, the Important filter Decreasing (ID) operation removes redundant filters with high functional similarity. Multiple ID operations are performed iteratively to gradually reduce the parameter size of the model without drastically breaking its performance. Experimental results show that compared with existing filter-level pruning methods, the compact model obtained by IDI has higher accuracy. IDI efficiently realizes the compression of deep neural networks.