Influence of Information Leakage in Analog Memory on Learning Kohonen Network on Silicon

The paper presents an influence of leakage effect observed in capacitive analog memories on learning process in hardware implemented Kohonen neural networks with MOS transistors used as switches connected with information holding capacitors. The learning results, i.e. variations (adaptations) of weight values, strongly depend on the transistor leakage currents. This is a cause for some quantization error associated with the weight adaptations during the network training. The unwanted leakage influence can be minimized in several ways discussed in this paper. As expected, the observed leakage influence on the memory storage time rises with an increase of temperature. This has been verified by means of computer simulations (Matlab, HSpice) as well as measurements of a prototyped Kohonen network chip (0.18 mu m CMOS process).

Published in:
Mixdes 2009: Proceedings Of The 16Th International Conference Mixed Design Of Integrated Circuits And Systems, 282-285
Presented at:
16th International Conference Mixed Design of Integrated Circuits and Systems, Lodz, POLAND, Jun 25-27, 2009
Piscataway, NJ, USA, Ieee Service Center

Note: The status of this file is: EPFL only

 Record created 2010-04-19, last modified 2018-03-17

Publisher's version:
Download fulltext

Rate this document:

Rate this document:
(Not yet reviewed)