ReCo1: A Fault resilient technique of Correlation Sensitive Stochastic Designs

TL;DR

This paper introduces ReCo, a fault-tolerant framework for correlation-sensitive stochastic logic circuits, significantly improving reliability and accuracy in error-prone stochastic systems, especially in critical applications like security and medical devices.

Contribution

The paper develops a novel ReCo framework for remodelling correlation in stochastic logic elements, enhancing fault tolerance without compromising overall circuit reliability.

Findings

Achieved a 92.80 SSIM in noisy image processing

ReCo reduces hardware area for correction blocks

Improves accuracy in correlation-sensitive stochastic circuits

Abstract

In stochastic circuits, major sources of error are correlation errors, soft errors and random fluctuation errors that affect the accuracy and reliability of the circuit. The soft error has the effect of changing the correlation status and in turn changes the probability of numbers leading to the erroneous output. This has serious impact on security and medical systems where highly accurate systems are required. We tackle this problem by introducing the fault-tolerant technique of correlation-sensitive stochastic logic circuits. We develop a framework of Remodelling Correlation(ReCo) for Stochastic Logic Elements; AND, XOR and OR for reliable operation. We present two variants of ReCo models in combinational circuits with contradictory requirements by stating two interesting case studies. The proposed technique selects logic elements and places correction blocks based on a priority-based rule that helps to converge to the desired MSE quickly requiring less hardware area. It is shown that this technique does not alter the reliability of the overall circuit. To demonstrate the practical effectiveness of the proposed framework, contrast stretch operation on a standard image in a noisy environment is studied. A high structural similarity index measure of 92.80 is observed for the output image with the proposed approach compared to the image (with error) 66.43.