2018/07/17 by Michael Lunglmayr, Lunglmayr, Michael, Daniel Wiesinger +3
Computer Science · #Error Correcting Code Techniques #FOS: Electrical engineering #Quantum Computing Algorithms and Architecture #Signal Processing (eess.SP) #Stochastic Gradient Optimization Techniques #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1807.06966
openalex publication_date 2018/07/17 · openalex created_date 2022/08/04 · openalex updated_date 2026/07/28
In stochastic computing (SC), a real-valued number is represented by a\nstochastic bit stream, encoding its value in the probability of obtaining a\none. This leads to a significantly lower hardware effort for various functions\nand provides a higher tolerance to errors (e.g., bit flips) compared to binary\nradix representation. The implementation of a stochastic max/min function is\nimportant for many areas where SC has been successfully applied, such as image\nprocessing or machine learning (e.g., max pooling in neural networks). In this\nwork, we propose a novel shift-register-based architecture for a stochastic\nmax/min function. We show that the proposed circuit has a significantly higher\naccuracy than state-of-the-art architectures at comparable hardware cost.\nMoreover, we analytically proof the correctness of the proposed circuit and\nprovide a new error analysis, based on the individual bits of the stochastic\nstreams. Interestingly, the analysis reveals that for a certain practical bit\nstream length a finite optimal shift register length exists and it allows to\ndetermine the optimal length.\n