I have been trying to train an MLP regression model.
The data has way more dimensions than samples.
The approach I took is first run PCA on it such that the amount of left over dimensions < samples.
What I seem to experience (using GridSearchCV) is that the 'best fit' results in a 'neural network' with just one layer and one node (activation is logistic). That does not feel ok, so I was wondering if there are any conclusions to take from this. Perhaps the data is not a good predictor as such? Or is PCA not the right approach? Or should I try something else than a MLPRegressor model?
All tips greatly appreciated.
Peter
