Table 1 Performance comparison of PNNs for benchmark tasks
Data set | Training method | ||||
|---|---|---|---|---|---|
MNIST | Fashion MNIST | Cifar-10 | |||
PNN based on photonics | Deep RC [sim.] (this study) | 98.96 % (w/o preprocessing) | 86.52% | 55.80% | Augmented DFA (with optoelectronic hardware) |
Deep RC [exp.] (this study) | 97.80% (w/o preprocessing) | 85.91% | 47.83% | ||
Diffractive photonic DNN | 96.6 % [exp.]34 (w/o preprocessing) | 84.6% [exp.]34 | 44.4% [exp.]69 | BP (on external standard computer) | |
On-chip photonic DNN | 95.3% [exp.]66 (w/o preprocessing) | Not reported | Not reported | ||
Large-scale Photonic RC | 97.15% [sim.]68 (w/o preprocessing) 98.90% [exp.]68 (with preprocessing) | Not reported | Not reported | Linear regression (on external standard computer) | |
On-chip Photonic RC | 91.3% [exp.]35 (w/o preprocessing) | 70.1% [exp.]35 | Not reported | ||
RC based on other physical dynamics | Spintronic RC | 87.6% [sim.]65 (w/o preprocessing) | Not reported | Not reported | Linear regression (on external standard computer) |
Memristor RC | 88.1% [exp.]64 (w/o preprocessing) | Not reported | Not reported | ||
Diffusive memristor RC | 83% [exp.]67 (w/o preprocessing) | Not reported | Not reported | ||
Self-organized nanowire RC | 90.4% [exp.]63 (w/o preprocessing) | Not reported | Not reported | ||
SOTA model on standard computer (PNNs not used) | 99.91% (Ensembled CNN70) | 95.99% (ResNet-11071) | 98.9% (Efficient Net72) | BP (all computations done by standard computer) | |
