Table 2.
Self-repair results for A-STDP enabled SNNs.
| Network description | Fault probability | Accuracy after fault injection (%) | Accuracy after weight normalization (%) | Accuracy after STDP re-training (%) | Accuracy after A-STDP re-training (%) | Accuracy gain from A-STDP |
|---|---|---|---|---|---|---|
| MNIST Dataset | ||||||
| 225 Excitatory neuronsBaseline accuracy = 89.53% | 50% | 75.41 ± 2.28 | 83.04 ± 0.49 | 76.69 ± 0.98 | 84.06 ± 0.70 | 1.02 |
| 60% | 70.72 ± 1.41 | 80.25 ± 0.70 | 73.85 ± 1.18 | 82.19 ± 0.28 | 1.95 | |
| 70% | 61.86 ± 2.03 | 76.06 ± 1.22 | 70.57 ± 0.48 | 79.39 ± 1.13 | 3.33 | |
| 80% | 30.42 ± 3.26 | 69.13 ± 0.85 | 67.42 ± 1.37 | 75.42 ± 0.48 | 6.29 | |
| 90% | 9.92 ± 0.11 | 56.69 ± 1.45 | 61.40 ± 1.38 | 65.57 ± 1.28 | 8.89 | |
| 400 Excitatory neuronsBaseline accuracy = 92.02% | 50% | 79.32 ± 2.57 | 85.56 ± 0.24 | 80.96 ± 1.24 | 87.16 ± 0.12 | 1.59 |
| 60% | 73.01 ± 2.15 | 82.61 ± 0.22 | 79.12 ± 1.28 | 85.17 ± 0.33 | 2.56 | |
| 70% | 61.20 ± 1.10 | 79.77 ± 0.61 | 77.51 ± 0.62 | 83.00 ± 0.40 | 3.22 | |
| 80% | 30.18 ± 2.60 | 73.08 ± 0.87 | 73.26 ± 1.16 | 78.68 ± 0.73 | 5.58 | |
| 90% | 9.80 ± 0.27 | 59.90 ± 1.16 | 67.80 ± 0.77 | 68.85 ± 0.48 | 8.95 | |
| Fashion-MNIST dataset | ||||||
| 400 Excitatory neuronsBaseline accuracy = 77.35% | 50% | 58.62 ± 1.24 | 73.85 ± 0.50 | 73.51 ± 0.30 | 75.88 ± 0.38 | 2.02 |
| 60% | 39.12 ± 1.19 | 71.85 ± 1.36 | 72.23 ± 0.60 | 75.16 ± 0.49 | 3.31 | |
| 70% | 16.61 ± 0.69 | 70.21 ± 0.44 | 70.63 ± 0.70 | 73.14 ± 0.44 | 2.93 | |
| 80% | 10.00 ± 0.22 | 66.32 ± 0.58 | 68.80 ± 0.47 | 70.82 ± 0.57 | 4.51 | |
| 90% | 10.04 ± 0.28 | 60.24 ± 0.86 | 63.92 ± 0.77 | 65.49 ± 0.40 | 5.25 | |