Table 4.
Comparison of previously published classification results on the MNIST dataset for spiking neural networks that are trained using supervised learning with a single, fully connected (non-convolutional) hidden layer and temporal encoding of input data. The second column provides the number of hidden neurons
| Publication | # Hidden | Test accuracy | Comments |
|---|---|---|---|
| This Work | 350 | 97.6(1)% | |
| Cramer et al.42 | 246 | 97.5(1)% | Downsampled to 16 by 16 pixels |
| Zenke and Vogels43 | 512 | 98.3(9)% | Including recurrent connections |
| Kheradpisheh and Masquelier30 | 400 | 97.4(2)% | |
| Comsa et al.28 | 340 | 97.9% (Max.) | Bias spikes at learned times |
| Göltz et al.27 | 350 | 97.5(1)% | |
| Mostafa29 | 800 | 97.55% | |
| Neftci et al.44 | 500 | 97.77% (Max.) | |
| Lee et al.45 | 800 | 98.71% (Max.) |