5.1 Answers#

N2V is a very peculiar deep-learning algorithm. It is a so-called self-supervised method. What is used as "ground-truth" in N2V?

The "ground-truth" is the actual pixel value of the noisy pixels that are masked during training.

Given your answer to the previous question, how can we interpret the N2V loss function value?

The loss function is the difference between the predicted pixel values and the original masked pixel values. Therefore, the loss will never go to 0 (unless the network is learning the identity function, which by design it cannot here). There must be a difference between the prediction and the original value, but its absolute value is meaningless.

What could happen to the validation loss curve if you have too little validation examples?

It could fluctuate massively if you have too little validation data. That's why the minimum of validation data in the notebook is 10% of the total amount of data.