Noise Tolerant probabilistic logic for statistical pattern recognition applications

Abstract

The new generation of knowledge-based applications requires a large amount of computing power with minimal energy consumption. This has aroused the interest in the non-conventional computing methods capable to implement complex functions in a very simple way and which in turn are inherently noise tolerant, as is the case of probabilistic or stochastic computing architectures. This work analyzes the robustness against noise of the Extended Stochastic Logic (ESL) encoding, a recently proposed probabilistic computing methodology. Furthermore, the capabilities of the ESL encoding to implement complex computational functions in the field of statistical pattern recognition, as is the case of a Bayesian classifier, are presented. The ESL noise-tolerance is analyzed and tested in a FPGA by injecting a wide range of noise levels. The noise-tolerance results are compared with the archived by conventional circuits, with and without fault-tolerant capabilities. The ESL outperforms the conventional Triple Modular Redundancy (TMR) solutions as is show in the experimental results.