Table 4.
The “Network in Network” (NiN) architecture with batch-normalization (BN) (Ioffe and Szegedy, 2015) used for CIFAR-100.
| Layer | Kernel size | Number of features | Stride | Non-linearity | Other |
|---|---|---|---|---|---|
| Conv-1 | 5 × 5 | 192 | 1 | ReLU | BN |
| MLPConv-1-1 | 1 × 1 | 160 | 1 | ReLU | BN |
| MLPConv-1-2 | 1 × 1 | 96 | 1 | ReLU | BN |
| MaxPool | 3 × 3 | 96 | 2 | Max | − |
| Conv-2 | 5 × 5 | 192 | 1 | ReLU | BN, Dropout (p = 0.5) |
| MLPConv-2-1 | 1 × 1 | 192 | 1 | ReLU | BN |
| MLPConv-2-2 | 1 × 1 | 192 | 1 | ReLU | BN |
| AveragePool-1 | 3 × 3 | 192 | 2 | − | − |
| Conv-3 | 5 × 5 | 192 | 1 | ReLU | BN, Dropout (p = 0.5) |
| MLPConv-3-1 | 1 × 1 | 192 | 1 | ReLU | BN |
| MLPConv-3-2 | 1 × 1 | 100 | 1 | ReLU | BN |
| AveragePool-2 | 8 × 8 | 100 | − | − | − |