There are a lot of things we should try!
For example:
- The NN might have an easier time learning
true_pt / pt_reconstructed, can you feed in the pt ratios and try to compute it as well? - We didn’t normalize the inputs at all. NNs train slightly more
easily if you feed them gaussian inputs. There are a few ways to
accomplish this.
- You could use pt ratios rather than raw pt, i.e. subjet pt / large jet pt, or (for the target) truth pt / reconstructed pt
- You can use functions like log or just linear scaling
And then, there are many things we did wrong!
- You should use a training, validation, and testing set, but for that we’d need a bit more time and data.
- You never implemented this in a real analysis! Booo. There’s more information in mlbnn