Table 1.
Comparison with other inference-latency reducing methods.
| Dataset | Method | Accuracy (decline compared with CNN) | Inference-latency (time steps) | Accelerative ratio |
|---|---|---|---|---|
| MNIST (Neil et al., 2016) | Sparse Coding | 98.00% | 631 | - |
| MNIST (Neil et al., 2016) | Activation Cost | 98.00% | 602 | - |
| MNIST (Neil et al., 2016) | Dropout | 98.00% | 641 | - |
| MNIST (Neil et al., 2016) | Dropout Learning Sched. | 98.00% | 602 | - |
| MNIST (Neil et al., 2016) | Stacked AE | 98.00% | 788 | - |
| MNIST (Avg 0b Analog) | Stopping criterion | 98.50% (0.06%) | 24 | 1.88X |
| MNIST (Avg 0b Poisson) | Stopping criterion | 98.48% (0.08%) | 27 | 1.48X |
| MNIST (Max 0b Analog) | Stopping criterion | 97.91% (0.74%) | 30 | 1.97X |
| MNIST (Avg BN Analog) | Stopping criterion | 98.73% (0.09%) | 70 | 1.39X |
| MNIST (Yang et al., 2020) | Conversion rule | 99.03% (0.08%) | 67 | 1.49X |
| CIFAR-10 (Avg 0b Analog) | Stopping criterion | 87.72% (0.23%) | 267 | 1.87X |
| CIFAR-10 (Avg 0b Analog) | Stopping criterion | 87.25% (0.70%) | 146 | 1.81X |
| CIFAR-10 (Yang et al., 2020) | Conversion rule | 80.03% (0.78%) | 245 | 1.63X |
The accelerative ratio in the table is compared with original model (Rueckauer et al., 2017) under the same accuracy.