@yamins81
If we want to fit models to behavioral data, I imagine we want to be able to optimize this objective jointly with correctly predicting the label of an image. I think this can work by just having 2 different cost layers, one for performance, and one for predicting the behavioral data with a square error cost.
If we don't have behavioral data for a given image, is there a way to turn off one of the cost layers (have it not backpropagate gradient for that image or batch) without having to modify cuda code?
I fear that restarting the training procedure for some batches will cause problems with the momentum of the weights, but maybe that will be ok.
@yamins81 If we want to fit models to behavioral data, I imagine we want to be able to optimize this objective jointly with correctly predicting the label of an image. I think this can work by just having 2 different cost layers, one for performance, and one for predicting the behavioral data with a square error cost.
If we don't have behavioral data for a given image, is there a way to turn off one of the cost layers (have it not backpropagate gradient for that image or batch) without having to modify cuda code?
I fear that restarting the training procedure for some batches will cause problems with the momentum of the weights, but maybe that will be ok.