When NNs don’t work, we have many choices:
- Fetch more data
- Add more layers to NN
- Try some new approaches in NN
- Train longer(increase the number of iterations)
- Change batch size
- Try regularization
- Chcek Bias Variance trade-off to avoid under and overfitting
- Use more GPUs for faster computation
Architecture:
classification: VGG, ResNet, DenseNet…
segmentation: FCN, Dilated Convolution, Mask RCNN
detection: Faster-RCNN, YOLO, SSD…
image generation: UNet, Dilated Convolution, DCGAN, WGAN
Loss curve:
Regularization:
Dropout, weight decay
