Table 2.
The comparison between state-of-the-art methods and the proposed spiking gating flow (SGF) network.
| Name | Type | Learning | Learning | Model information | Training cost | Accuracy | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| method | style | Size | Diff(×) | OPs | Diff(×) | Epoch | T/I ratio | |||
| Reservoir CSNN (George et al., 2020) | SNN | STDP | Offline | 3.17 MB | 88.7↑ | - | - | 3.8:1 | 65.0% | |
| Heterogeneity Network (Perez-Nieves et al., 2021) | SNN | SGD | Offline | 125 KB | 3.4↑ | - | - | 3.8:1 | 82.1% | |
| SLAYER (Shrestha and Orchard, 2018) | SNN | BP | Offline | 1034.8KB | 28.3↑ | 79.8M | 9.6↑ | 739 | 3.8:1 | 93.64% |
| SCRNN (Xing et al., 2020) | ANN2SNN | BPTT | Offline | 732.34KB | 20.0↑ | 81.91 M | 9.9↑ | 100 | 4.1:1 | 96.59% |
| Converted SNN (Kugele et al., 2020) | ANN2SNN | BP | Offline | 500KB | 13.7↑ | 651 M | 78.7↑ | 10 | 3.8:1 | 96.97% |
| ConvNet (Amir et al., 2017) | DNN2SNN | BP | Offline | 16.3 MB | 456↑ | 946.82 M | 114↑ | 250 | 3.8:1 | 96.5% |
| PointNet++ (Qi et al., 2017) | DNN2SNN | BP | Offline | 3.50MB | 98↑ | 440.0 M | 53.2↑ | 250 | 3.8:1 | 97.08% |
| This work | SNN | SGF | Online | 36.58 KB | 8.27 M | 1 | 1.5:1 | 87.5% | ||
“-” Indicates the data can not be calculated or not mentioned in the corresponding paper. Bold indicate the results in our method. ↑ Indicates that our method has improvement compared to the related works.