Table 3. Comparison with other models containing baseline basic blocks on CIFAR-10 dataset.
| Model | W/Ab | Down samplingc | Top-1 (%)d | Storage (Mbits) | FLOPS (×107) |
|---|---|---|---|---|---|
| ResNet-18 (He et al., 2016)a | 32/32 | FPconv | 93.02 | 358 | 58.6 |
| XNOR-Net-18 (Rastegari et al., 2016) | 1/1 | Bconv | 89.83 | 12.9 | 1.41 |
| Bi-Real-Net-18 (Liu et al., 2018) | 1/1 | FPconv | 89.30 | 18.2 | 3.82 |
| AresB-18 | 1/1 | No conv | 91.90 | 12.7 | 1.36 |
Notes:
a
A ResNet-18 model contains eight basic blocks.
b
Terms W and A denote the precision of target weights and activation.
c
Prefix FP and B mean the full-precision and binarized 1 × 1 convolutions, respectively.
d
Top-1 accuracy indicates the final Top-1 test accuracy on the CIFAR-10 dataset.