Stochastic Associative Processor Operated by Random Voltages

Michihito UEDA  Ichiro YAMASHITA  Kiyoyuki MORITA  Kentaro SETSUNE  

Publication
IEICE TRANSACTIONS on Electronics   Vol.E90-C   No.5   pp.1027-1034
Publication Date: 2007/05/01
Online ISSN: 1745-1353
DOI: 10.1093/ietele/e90-c.5.1027
Print ISSN: 0916-8516
Type of Manuscript: Special Section PAPER (Special Section on Fundamentals and Applications of Advanced Semiconductor Devices)
Category: LSI Applications
Keyword: 
stochastic computing,  random variables,  SPICE simulator,  Manhattan distance,  and vector matching,  

Full Text: PDF(467.6KB)>>
Buy this Article




Summary: 
The latest LSIs still lack performance in pattern matching and picture recognition. Living organisms, on the other hand, devote very little energy to processing of this type, suggesting that they operate according to a fundamentally different concept. There is a notable difference between the two types of processing: the most similar pattern is always chosen by the conventional digital pattern matching process, whereas the choice made by an organism is not always the same: both the most similar patterns and other similar patterns are also chosen stochastically. To realize processing of this latter type, we examined a calculation method for stochastically selecting memorized patterns that show greater similar to the input pattern. Specifically, by the use of a random voltage sequence, we executed stochastic calculation and examined to what extent the accuracy of the solution is improved by increasing the number of random voltage sequences. Although calculation of the Manhattan distance cannot be realized by simply applying stochastic computing, it can be done stochastically by inputting the same random voltage sequence to two modules synchronously. We also found that the accuracy of the solution is improved by increasing the number of random voltage sequences. This processor operates so efficiently that the power consumption for calculation does not increase in proportion to the number of memorized vector elements. This characteristic is equivalent to a higher accuracy being obtained by a smaller number of random voltage sequences: a very promising characteristic of a stochastic associative processor.