Table 3 Performance of a spiking network on the split-MNIST task in different settings. The table shows the mean and standard deviation of the individual task accuracies and the mean accuracy across tasks over 5 runs after sequential training.
From: Probabilistic metaplasticity for continual learning with memristors in spiking networks
Task | Baseline | Random consolidation | Decaying probabilistic plasticity | Probabilistic metaplasticity | ||
|---|---|---|---|---|---|---|
(Factor=2) | (Factor=5) | (Factor=10) | ||||
Task 1 | 44.31 ± 3.80 | 48.01 ± 2.77 | 47.04 ± 1.63 | 50.11 ± 1.89 | 73.58 ± 2.80 | 84.95 ± 3.42 |
Task 2 | 53.67 ± 1.42 | 57.14 ± 0.61 | 56.39 ± 0.62 | 94.31 ± 0.27 | 94.76 ± 0.43 | 90.20 ± 1.96 |
Task 3 | 8.63 ± 1.04 | 17.73 ± 1.09 | 22.42 ± 1.53 | 90.51 ± 0.40 | 84.59 ± 0.59 | 72.06 ± 2.60 |
Task 4 | 88.32 ± 1.24 | 96.24 ± 0.55 | 96.68 ± 0.44 | 93.59 ± 0.91 | 82.72 ± 1.39 | 94.73 ± 0.94 |
Task 5 | 97.52 ± 0.32 | 95.73 ± 0.62 | 95.17 ± 0.38 | 48.86 ± 1.66 | 40.44 ± 1.16 | 76.54 ± 2.60 |
Mean | 58.49 ± 0.73 | 62.97 ± 0.56 | 63.54 ± 0.36 | 75.47 ± 0.32 | 75.22 ± 0.79 | 83.70 ± 0.78 |
