How to penalize a category error more than others categories during training (SoftMaxWithLoss layer)

[Hermann Hesse]

Hi all,

is there any way to penalize more a category error during training the network? For example, accumulating two times more than standard for a category error when computing the loss.

I'm using SoftMaxWith(cross-entropy)Loss layer, but I can use another cost function if necessary. I guess that i should edit the softmax_loss_layer.cu file, but my CUDA level is quite low.

Thanks so much.

[Hermann Hesse]

Anyone please?