For example, I have layers that are pretrained. But while predicted, the loss is very high. But not because of pre-trained layers. Because of not pretrained layers. Will every layer be affected by backprop the same?
1 Answer
This depends on how you configure the training process:
You can, for instance, freeze the pretrained layers; this implies that only the not pretrained layers will be updated.
You can also set different learning rates to different layers, so that the pretrained layers are assigned a very small learning rate that allows them to be updated but not too fast.
Therefore, backpropagation is the same for all layers but the weight update strategy can be different.