Abstract

Filters in a Convolutional Neural Network (CNN) contain model parameterslearned from enormous amounts of data. In this paper, we suggest to decomposeconvolutional filters in CNN as a truncated expansion with pre-fixed bases,namely the Decomposed Convolutional Filters network (DCFNet), where theexpansion coefficients remain learned from data. Such a structure not onlyreduces the number of trainable parameters and computation, but also imposesfilter regularity by bases truncation. Through extensive experiments, weconsistently observe that DCFNet maintains accuracy for image classificationtasks with a significant reduction of model parameters, particularly withFourier-Bessel (FB) bases, and even with random bases. Theoretically, weanalyze the representation stability of DCFNet with respect to inputvariations, and prove representation stability under generic assumptions on theexpansion coefficients. The analysis is consistent with the empiricalobservations.

 

Introduction (beta)

None

 

Conclusion (beta)

None