Functionality-Oriented Convolutional Filter Pruning

Zhuwei Qin (George Mason University), Fuxun Yu (George Mason University), Chenchen Liu (Clarkson University), Xiang Chen (George Mason University)

Abstract
The sophisticated structure of Convolutional Neural Network (CNN) models allows for outstanding performance, but at the cost of intensive computation load. To reduce this cost, many model compression works have been proposed to eliminate insignificant model structures, such as pruning the convolutional filters which have smaller absolute weights. However, most of these works merely depend on quantitative significance ranking without qualitative filter functionality interpretation or thorough model structure analysis, resulting in considerable model retraining cost. Different from previous works, we interpret the functionalities of the convolutional filters and identify the model structural redundancy as repetitive filters with similar feature preferences. In this paper, we proposed a functionality-oriented filter pruning method, which can precisely remove the redundant filters without compromising the model functionality integrity and accuracy performance. Experiments with multiple CNN models and databases testified the unreliability of conventional weight-ranking based filter pruning methods, and demonstrate our method’s advantages in terms of computation load reduction (at most 68.88% FLOPs), accuracy retaining (<0.34% accuracy drop), and expected retraining independence.

DOI
10.5244/C.33.229
https://dx.doi.org/10.5244/C.33.229

Files
Paper (PDF)

BibTeX