Table 1.
Validation accuracy and backward pass speedup per epoch for different datasets.
| Validation accuracy [%] | Mean backward pass speedup compared to BPTT | |||
|---|---|---|---|---|
| EXODUS | SLAYER | EXODUS | SLAYER | |
| CIFAR10-DVS | 72.28 ± 0.13 | 71.53 ± 0.18 | – | – |
| DVS gesture (5 layers) | 92.8 ± 2.2 | 87.8 ± 3.0 | 16.57 × | 14.09 × |
| DVS gesture (8 layers) | 94.54 ± 0.8 | 93.64 ± 0.49 | – | – |
| SHD | 78.01 ± 0.2 | 70.58 ± 1.9 | 9.1 × | 11.22 × |
| SSC | 55.41 ± 0.4 | 40.1 ± 0.8 | 17.3 × | 14.59 × |
| Average for 3 datasets | 14.3 × | 13.3 × | ||
Validation accuracy is mean and standard deviation of maximum accuracies across three runs with a gradient scale of 1. Backward pass speedup is measured across three epochs on a NVIDIA GeForce 1,080 Ti. Bold indicate best value for each category (accuracy, speedup) and dataset.