Construction of CNN Template via Learning with Random Weight Change Algorithm
Conference: CNNA 2016 - 15th International Workshop on Cellular Nanoscale Networks and their Applications
08/23/2016 - 08/25/2016 at Dresden, Deutschland
Proceedings: CNNA 2016
Pages: 2Language: englishTyp: PDFPersonal VDE Members are entitled to a 10% discount on this title
Adhikari, Shyam P.; Yang, Changju; Kim, Hyongsuk (Division of Electronics Engineering, Chonbuk National University, Jeonju, South Korea)
Chua, Leon O. (Dept. of Electrical Engineering and Computer Sciences, University of California, Berkeley, CA94720, USA)
A construction method of CNN cloning template is proposed via learning with Random Weight Change algorithm. For this learning, a target image for each input image is prepared via a sketch or any kind of image processing technique. After an initial template setting, a vector of randomly generated small values is added to the CNN template and tested upon input-target image pair. If the learning error is smaller than the one before updating, the template is taken as the one for learning in the next iteration and the same vector of random values is added to this template again. Otherwise, a new vector for updating the template is regenerated. Since the circuit implementation of this algorithm is feasible, this method is well suited for hardware-based CNN template learning.