Table 4 Evaluation of probabilistic metaplasticity and neural activity-dependent weight consolidation models on the split-MNIST and split-Fashion MNIST tasks in the domain-IL scenario.
From: Probabilistic metaplasticity for continual learning with memristors in spiking networks
Model | Split-MNIST | Split-FMNIST | Normalized memory overhead |
|---|---|---|---|
Probabilistic metaplasticity | |||
Individual m | 83.70 ± 0.78 | 93.23 ± 0.14 | \(\sim\) 0.3 MB |
Neuron-shared m | 82.07 ± 0.84 | 92.72 ± 0.30 | \(\sim\) 0.4 kB |
Layer-shared m | 79.74 ± 1.21 | 92.01 ± 0.49 | 4 Bytes |
NAI44 | 68.35 ± 1.34 | 68.82 ± 1.15 | \(\sim\) 0.8 kB |
NEO42 | 78.14 ± 2.23 | 86.82 ± 0.60 | \(\sim\) 0.3 MB |
