کنفرانس مهندسی برق ایران

صفحه اصلی / بیست و نهمین کنفرانس مهندسی برق ایران

Low Cost Implementation of Neural Networks Based on Stochastic Computing

نویسندگان :

Hadi Jahanirad¹ Ahmad Menbari²

1- دانشگاه کردستان 2- دانشگاه کردستان

کلمات کلیدی :

multi-layer perceptron, random number generator, stochastic computation, probability estimator

چکیده :

Implementation of a multi-layer perceptron (MLP) by conventional circuits requires a very complex a hardware demanding circuit design. Recently, stochastic computation elements (SC) show the great potential to implement MLP with a low hardware complexity. The conversion between binary numbers and stochastic sequences has dramatic effect on accuracy and speed of an MLP implementation. In this paper, two novel architectures for digital to probability converter (DPC) and probability to digital converter (PDC) are proposed to improve the accuracy and the precision of an MLP implantation. The novel DPC performs the conversion by higher precision than the previous ones. The stochastic sequences converts to their corresponding binary numbers by 100% precision using the proposed PDC. Furthermore, the required number of clocks to generate the binary numbers are lower than the best method (~0.1 by average) and are equal for all of the numbers.