Is there a learning process in the testing scenario, like in training?

Question

I am a beginner and I've been confused by my lecturer's request to show a loss function graphically (using tensorboard) for a testing dataset.

I'm currently working to implement single shot detector (SSD) for pavement distress detection in video. I've been able to use tensorboard to show total loss graphics for the training dataset.

As far as I know, the loss function is more related to the training dataset because of the learning process and my lecturer says "because it is the same model so the testing should have a learning process as well therefore there is a loss function whether it is training or testing".

So, here is my question:

1. Is there really a learning process in a testing scenario like in training?
2. If there is, how to calculate the loss function in testing scenario?
3. If there isn't, how to explain why there is no learning process (or loss function) in testing scenario even though it is the same model?

Answer 1

A loss function is used to quantify the quality of the prediction.

The training, the loss is used to adapt parameters by backpropagation. (Propagating the loss through the layers).

In testing, you do not backpropagate the loss, you nonetheless want to report it as you test your solution / the current state of the network.

Thus you might report the loss in both cases.