Table 2.
Different structures for implementing first convolutional layers.
| Name | Convolution | Bias | Nonlinearity | Norm | Pooling | Nonlinearity | More info |
|---|---|---|---|---|---|---|---|
| AlexNet | Conv2D(11*11) and stride of 4 | Bias | ReLU | LRN | Max-pool(3*3) and stride of 2 | – | – |
| Trans-ONN-vert | Wave-optics Conv | – | SA | – | Wave-optics Motion-pool (by vertical mask) | Sqnl and DS(4*4) | – |
| Trans-ONN-vert-GAP | Wave-optics Conv | – | SA | – | Wave-optics Motion-pool (by vertical mask) | Sqnl and DS(4*4) | Considering GAP instead of FC6 and FC7 |
| Trans-ONN-horiz | Wave-optics Conv | – | SA | – | Wave-optics Motion-pool (by horizontal mask) | Sqnl and DS(4*4) | – |
| Trans-ONN -horiz-GAP | Wave-optics Conv | – | SA | – | Wave-optics Motion-pool (by horizontal mask) | Sqnl and DS(4*4) | Considering GAP instead of FC6 and FC7 |
| Trans-ONN -cascaded | Wave-optics Conv | – | SA | – | Wave-optics Motion-pool (2 cascaded 4f systems) | Sqnl and DS(4*4) | – |
| Trans-ONN -cascaded-GAP | Wave-optics Conv | – | SA | – | Wave-optics Motion-pool (2 cascaded 4f systems) | Sqnl and DS(4*4) | Considering GAP instead of FC6 and FC7 |