Table 1.
Accuracy of spiking neural network (SNN) trained using BlocTrain and end-to-end spike-based backpropagation through time (BPTT) methods, and SNN/analog neural network (ANN) trained using only the local losses, on the CIFAR-10 dataset.
| Model | Training method | Dataset size | %Accuracy |
|---|---|---|---|
| CIFARNet w/ 7 layers (Wu et al., 2019) | End-to-end STBP (Wu et al., 2018) | 50,000 | 90.53 |
| ResNet-9 (Lee et al., 2020) | End-to-end Spike BP | 50,000 | 90.35 |
| SNN w/ 8 layers (Thiele et al., 2020) | End-to-end ANN-based SpikeGrad | 50,000 | 89.72 |
| ResNet-11 (Ledinauskas et al., 2020) | End-to-end Spike BP | 50,000 | 90.2 |
| VGG-16 (Rathi et al., 2020) | ANN-SNN and end-to-end STDB | 50,000 | 91.13 |
| VGG-16 (Zhou et al., 2020) | Direct end-to-end BP | 50,000 | 92.68 |
| SNN w/ 4 layers (Panda and Roy, 2016) | Local AutoEncoder | 50,000 | 70.16 |
| ANN w/ 10 layers (Mostafa et al., 2018) | Local training | 50,000 | ~83 |
| ResNet-9 (our work) | BlocTrain | 40,000 | 86.4 |
| ResNet-9 (our work) | BlocTrain-base | 50,000 | 88.31 |
The bold values are used to highlight the results reported in this work over prior works.