Fast Classifier

Fast Algorithm for GMM-Based Pattern Classifier


  • Gussian distribution is popularly used in statistical pattern classification problems.
    Not suitabe for modeling a multi-modal distribution.
  • Gussian mixture model (GMM) can approximate a multi-modal distribution and be an alternative.
    Higher computational costs are not preferable.
  • Statistical pattern classification problems often meet a situation that comparison between probabilities is obvious and redundant.
  • In this work, an efficient implemetation of the exponential function is proposed for GMM-based pattern classification.
    • A hardware friendly algorithm is obtained.
    • Evaluation on programmable DSP shows the significance.
    • Adaptive control of computational precision is achieved to reduce the redundant operations.


A comparison of the exponential function’s evaluation is replaced by a comparison of the intervals based on the following inequality:

    \[ 2^{-(\lfloor z\log_2 e\rfloor+1)}<\exp(-z)\leq 2^{-\lfloor z\log_2 e\rfloor}, \]

where z>0. Since \log_2 e is constant (1.442695040888963…), the interval calculation is achieved only by constant scaling of positive variable z, flooring and bit shifts.

Journal Paper

  1. Hidenori Watanabe and Shogo Muramatsu: Fast Algorithm and Efficient Implementation of GMM-Based Pattern Classifiers, Journal of Signal Processing Systems, Springer, Volume 63, Number 1, April 2011 , pp. 107-116(10), DOI: 10.1007/s11265-009-0439-z, Apr. 2011. (Online)


  1. Shogo Muramatsu and Hidenori Watanabe: Fast Algorithm for GMM-Based Pattern Classifier, Proc. of 2009 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP2009), pp.633-636, Taipei, Apr. 2009.


  1. Identification Device, Identification Method, and Identification Processing Program,Shogo Muramatsu, Hidenori Watanabe
    ( 8,321,368,Nov. 27, 2012) ,USA