Performance of Post-Training Two-Bits Uniform and Layer-Wise Uniform Quantization for MNIST Dataset from the Perspective of Support Region Choice

Table 1

SQNR and model accuracy for application of different two-bits UQ designs.

Two-bits UQ, accuracy (FP32) = 98.1%
xmin = -7.063787, xmax = 4.8371024; xmax^(H) = 1.9605, xmax^(J) = 2.1748	Case 1 ℜg	Case 2 ℜg	Case 3 ℜg	Case 4 ℜg
	[−xmax, xmax]	[xmin, −xmin]	[−xmax^(H), xmax^(H)]	[−xmax^(J), xmax^(J)]
SQNRex^UQ (dB)	2.8821	−1.2402	8.7676	8.7639
SQNRth^UQ (dB)	1.9360	−2.0066	6.9787	7.0707
Accuracy (%)	96.97	94.58	96.34	96.74
Within ℜg (%)	99.988	100	94.787	96.691