eduzhai > Applied Sciences > Engineering >

Exploring the parameter reusability of CNN

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: In recent times, using small data to train networks has become a hot topic inthe field of deep learning. Reusing pre-trained parameters is one of the mostimportant strategies to address the issue of semi-supervised and transferlearning. However, the fundamental reason for the success of these methods isstill unclear. In this paper, we propose a solution that can not only judgewhether a given network is reusable or not based on the performance of reusingconvolution kernels but also judge which layers parameters of the givennetwork can be reused, based on the performance of reusing correspondingparameters and, ultimately, judge whether those parameters are reusable or notin a target task based on the root mean square error (RMSE) of thecorresponding convolution kernels. Specifically, we define that the success ofa CNN s parameter reuse depends upon two conditions: first, the network is areusable network; and second, the RMSE between the convolution kernels from thesource domain and target domain is small enough. The experimental resultsdemonstrate that the performance of reused parameters applied to target tasks,when these conditions are met, is significantly improved.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×