Application and Evaluation of Quantization for Narrow Bit-width Resampling of Sequential Monte Carlo
Hiroki Nishimoto, Renyuan Zhang, Yasuhiko Nakashima · 2022
As a preliminary step to reduce the resampling circuit size of the Sequence Monte Carlo (SMC) method, an evaluation of convergence feasibility by preprocessing is presented in this paper. SMC is a powerful numerical algorithm widely applied to Bayesian estimation and Kalman filter. However, SMC is computationally expensive and difficult to apply to huge amounts of data. Several previous studies have already reported methods to accelerate SMC using GPGPUs and FPGAs. In this study, we attempted to find a threshold that converges even with a narrow register width by quantizing the weights used for resampling through various pre-processing methods. Experimental results showed that with pre-processing, 8-bit unsigned integer can achieve the same performance as 32-bit floating point.