search

chengsoonong / crowdastro

Cross-identification of radio objects and host galaxies by applying machine learning on crowdsourced training labels.
MIT License
13 stars 1 forks source link

Investigate colour/intensity normalisation #225

Open chengsoonong opened 7 years ago

MatthewJA commented 7 years ago

This could be arbitrarily hard; could we mark it wontfix?

chengsoonong commented 7 years ago

My thought for this issue is to do some non-linear squashing (something like a sigmoid) and that should do the most of the required job.

I recall @jbanfield suggesting something like asinh.

MatthewJA commented 7 years ago

Alright, let's use asinh; I recall seeing a source saying that asinh has been shown to work well on inputs to these things.

MatthewJA commented 7 years ago

I can't seem to get fitting logistic regression to actually terminate with an asinh stretch (working on it...) but for random forests the performance is improved a fair bit with such a stretch.