From: Neural network training with limited precision and asymmetric exponent
Mantissa’s bit count | 4-bit regular exponent (%) | 3-bit asymmetric exponent (%) | Accuracy improvement (%) |
---|---|---|---|
4 | 10.10 | 75.89 | 65.79 |
5 | 10.10 | 85.97 | 75.87 |
6 | 39.21 | 89.31 | 50.10 |
7 | 66.72 | 90.27 | 23.55 |
8 | 87.28 | 95.01 | 7.73 |
9 | 91.29 | 94.07 | 2.78 |
10 | 94.12 | 97.13 | 3.01 |
11 | 95.88 | 97.76 | 1.88 |