Table 1.
Hybrid optoelectronic CNN performance for grayscale CIFAR-10 classification with various models. Classification accuracies of simulated models are the average of five trials. Standard deviations (std. dev.) are computed for the simulated models. Learned parameters and FLOPs are split into those for the optical and electronic parts of the network, when relevant.
| Method | Test accuracy | Learned parameters | FLOPs |
|---|---|---|---|
| mean ± std. dev. | optical/electronic | optical/electronic | |
| Simulation: | |||
| fully connected (FC) only | 29.8 ± 0.5% | −/10,250 | −/20,480 |
| digital conv > ReLU > FC, unconstrained | 51.9 ± 1.3% | −/82,586 | −/1,490,954 |
| digital conv > ReLU > FC, nonnegative | 36.3 ± 0.5% | −/82,586 | −/1,490,954 |
| digital conv > ReLU > FC, pseudonegative | 51.8 ± 0.6% | −/83,234 | −/2,818,058 |
| optical conv > ReLU > FC, pseudonegative | 51.0 ± 1.4% | 104,976/81,938 | 3,779,136/180,234 |
| Physical experiment: | |||
| optical conv > ReLU > FC, pseudonegative | 44.4% | −/81,938 | −/180,234 |