Comparison with ML approach

[VForiel]

Another way of finding the good shifts to introduce is to generate a dataset containing series of kernel nuller input and outputs and the perfects shifts that should be injected and then train a dense neural network to find out the good shifts.

A possible advantage of this method is that it can avoid any kernel swapping or kernel reversing (as discussed in #53 ) de to the supervised nature of the training (we build the dataset only with direct solutions)

A drawback is the the output of such neural network is a random variable statistically distributed around the output we are searching for. However, the solution can be used as initial guess to run the classical calibration algorithm, which can converge to the nearest solution without having to scan the entire parameter space.