On Different Criteria for Optimizing the Two-bit Uniform Quantizer
Jelena Nikolić, Zoran Perić, Stefan S. Tomić, Danijela Aleksić · 2022
In this paper, we address the problem of determining the most influential parameter of the two-bit uniform quantizer, i.e. the support region threshold ($x$max), according to different optimization criteria. We analyze the dependences of the quantized neural network (QNN) model quality indicators on$\boldsymbol{x_{\max}}$, for the case of applying two-bit uniform quantization of weights in the post-training phase for the MNIST dataset. In addition to the theoretical quality indicator of quantized signal for the Laplacian distribution - theoretical SQNR, our quality indicators are also the accuracy of QNN and the experimentally determined SQNR for the Laplacian-like weight distribution in the three-layer fully connected neural network. The goal is to determine the results of optimizing these quality indicators per Xmaxfor the considered research framework and to make their comparison for setting a good foundation for future research.