Hi Sangmin,
I understand you want to define a loss function that requires ground truth data, intermediate network activations, and the final network output as input.
To implement such a loss function, you need to use the custom training loop approach instead of using trainNetwork. You can see an example here: Train Network Using Custom Training Loop.
If you look at the modelGradients function in that example, you'll be able to extract intermediate activations from the dlnetwork in the call to forward using the 'Outputs' name-value pair. And then you can formulate your own loss function without needing a regression layer.
There are already several typical loss functions available that you can make use of in your custom code, if applicable.
I hope this helps.
