A Latch-based Stochastic Number Generator for Stochastic Computing of Extended Naive Bayesian Network

Stochastic computing (SC) employing probability bitstreams is a low-cost and error-tolerant computing paradigm applied in Bayesian networks, etc. A stochastic number generator (SNG) is utilized to generate the bitstreams with specified probability. Instead of the conventional comparator and linear feedback shift register-based SNG, in this work, we propose 1) a latch-based stochastic number generator (L-SNG) that directly generates period-less true random bitstream with a specified probability of ones; 2) an extended naive Bayesian network that enhances classification accuracy by introducing evidence-to-evidence correlation factors in a conventional naive Bayesian network. Experimental results on a 130-nm CMOS test chip of L-SNG with the extended Bayesian network model using Sloth dataset demonstrate: 1) a 10% improvement in classification accuracy compared with the naive Bayesian network; 2) equivalent accuracy of 90% as in numerical computing using only 256-bit bitstreams; 3) high robustness with the same accuracy under conditions ranging from 0.7 V to 0.9 V and -20 degrees C to 100 degrees C.