Skip to main content
. 2021 Dec 20;23(12):1699. doi: 10.3390/e23121699

Table 2.

SQNR and QNN accuracy for MLP trained on the MNIST dataset with the application of different three-bit UQ designs.

wmin = −7.063787 wmax = 4.8371024
xmax[H] = 2.9408 xmax[J] = 2.9236
Choice 1 ℜg
[−wmax, wmax]
Choice 2 ℜg
[wmin, −wmin]
Choice 3 ℜg
[−xmax[H], xmax[H]]
Choice 4 ℜg
[−xmax[J], xmax[J]]
SQNRexUQ(dB) 9.2302 5.9005 12.9455 12.9766
SQNRthUQ(dB) 8.6901 5.1273 11.4414 11.4419
Accuracy (%) 97.85 97.85 97.66 97.65
Within ℜg (%) 99.988 100 99.403 99.381