Can I choose network for backpropagation?

[胡展維]

Hi ! I have my network which output three loss:

layer {
  name: "loss_1"
  type: "EuclideanLoss"
  bottom: "fc10_1"
  bottom: "fc_7"
  top: "loss_1"
}


layer {
  name: "loss_2"
  type: "EuclideanLoss"
  bottom: "fc10_2"
  bottom: "fc_7"
  top: "loss_2"
}


layer {
  name: "loss_3"
  type: "EuclideanLoss"
  bottom: "fc10_3"
  bottom: "fc_7"
  top: "loss_3"
}

and  I want to choose the minimum loss to be my output, after that, I want to backpropagate the network with minimum loss "ONLY" (means I don't want to backpropagate all the networks) !

Does anyone have some advice?

Thanks a lot !!

[Nam Vo]

Try this: MIN pooling and make it your loss function

[Jan C Peters]

Yes, MIN pooling could work. Do not forget setting the loss_weight parameters accordingly (set them to zero for the three loss layers above and to one for the min-pool layer).

Jan