- Weight decay, lambda = 0.001
- SGD with momentum = 0.9
- Learning rate gamma = [0.25, 0.1, 0.05, 0.01], 0.1 for now
- Adapt learning rate at [200, 250, 300] by multiplying by 0.1
- 350 epochs
- Train all NNs with CIFAR-10
- Visualize first 3 layers
| Model | Error (%) | Time (mins) |
|---|---|---|
| Model A | 15.16% | |
| Strided-CNN-A | 18.23% | |
| ConvPool-CNN-A | 11.44% | |
| ALL-CNN-A | 12.65% | 642.76 |
| Model B | 13.52% | |
| Strided-CNN-B | 15.56% | |
| ConvPool-CNN-B | 11.84% | 879.25 |
| ALL-CNN-B | 11.84% | |
| Model C | 11.11% | |
| Strided-CNN-C | 12.91% | |
| ConvPool-CNN-C | 11.09% | |
| ALL-CNN-C | 11.17% |