I'm doing a particle physics analysis (jet energy regression) by means deep neural network (in keras with tensorflow backend). I have several features (mostly kinematic variables). I trained my model on the HH->bbbb samples (Di-Higgs decaying to 2 pairs of b and anti-b quarks). I normalized this dataset to zero-mean unit variance (z-score normalization). Now, I want to predict using a different sample (HH->2b2g, di-Higgs decaying to a pair of b quarks and two photons). When predicting this dataset, should I normalize it based on the HH->bbbb statistics? When I try to do it, it doesn't predict well, even giving me negative values of pT (transverse momentum (pT) should be > 0). Should I normalize HH->2b2g samples based on its own statistics instead?
What you did is in principle correct, i.e. use the HH->bbb statistics to normalize the inputs of your new samples. The situation that you are facing may indicate that your new samples come from a very differente distribution compared to the one you used to train your model. You could possibly try to work out some maths in such a way that your model predicts a constrained output within a known range [min output value, max output value]. In this sense, if your initial model is predicting a value < 0, the modified one should predict 0.
Daniel Urda Thanks. By the way, If I did normalize my new inputs, how should I renormalize it? Should I do the inverse_transform of my HHbbbb statistics?
Exactly, if you applied the z-score normalization, then just calculate the inverse transformation using your HHbbb statistics (multiplying predictions by the standard deviation and then adding the mean).
- Badges
- Science topic