I was reading the tutorial "Statistical Significance Test for comparing ML algorithms", where it suggests to use k-fold and apply the appropriate statistical test.
Suppose that I have a train set and a test set and two deep learning neural networks in a regression problem. Since the training in deep learning takes lots of time, the test procedure using k-fold would be very costly regarding computation resources.
I was wondering if it is possible to apply the statistical tests only in the results of the test set.
For example, if the test set has 1000 samples, the two neural networks would be trained in the same training set and then apply Wilcoxon in the results of the two networks over the 1000 points of the test set, is that correct? Or do I always need to perform the k-fold? Besides can I use the MSE in each test point and compare those results?