Abstract
Stochastic computing (SC) is a probabilistic-based processing methodology that has emerged as an energy-efficient solution for implementing image processing and deep learning in hardware. The core of these systems relies on the selection of appropriate Random Number Generators (RNGs) to guarantee an acceptable accuracy. In this work, we demonstrate that classical Linear Feedback Shift Registers (LFSR) can be efficiently used for correlation-sensitive circuits if an appropriate seed selection is followed. For this purpose, we implement some basic SC operations along with a real image processing application, an edge detection circuit. Compared with the literature, the results show that the use of a single LFSR architecture with an appropriate seeding has the best accuracy. Compared to the second best method (Sobol) for 8-bit precision, our work performs 7.3 times better for the quadratic function; a 1.5 improvement factor is observed for the scaled addition; a 1.1 improvement for the multiplication; and a 1.3 factor for edge detection. Finally, we supply the polynomials and seeds that must be employed for different use cases, allowing the SC circuit designer to have a solid base for generating reliable bit-streams.
Highlights
Stochastic computing (SC) has emerged as a possible solution for Neural Network hardware implementation [1,2] and as a way to accelerate the computation in different applications such as image processing [3] or Deep Learning (DL) for inference [4,5] and training [6,7]
We demonstrate that the Linear Feedback Shift Registers (LFSR) may achieve low autocorrelation behavior when isolated properly, something that has been totally overlooked in the literature up to now
We have presented a solid base for the LFSR as the best Random Number Generators (RNGs) circuit in the SC domain, if computed for a complete sequence period
Summary
Stochastic computing (SC) has emerged as a possible solution for Neural Network hardware implementation [1,2] and as a way to accelerate the computation in different applications such as image processing [3] or Deep Learning (DL) for inference [4,5] and training [6,7]. The randomness quality can significantly affect the precision of those operations requiring non-correlated signals, as in the case of stochastic multiplication, for instance For these reasons, finding the best RNG in terms of area and low correlation is a major concern to address when designing real SC applications. Despite the fact that some solutions have been produced, none of them guarantees good accuracy and a small footprint area at the same time for correlation-sensitive circuits such as the SC quadratic function, the scaled addition, and the multiplication These circuits are the driving force in the SC realm, and high demand applications such as image processing or DL employ them. We explore in more detail how the LFSR circuit could be better exploited as a RNG source in SC by making a careful selection of the seeds employed, with the purpose of finding the best BS generator technique for different application requirements. We provide the seeds that must be employed when using LFSRs for different use cases, offering SC designers a direct RNG setting
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
