A bespoke loss function that optimised detection at fixed false alarm

hagabbar / matching_cnn_paper

Matching matched filtering with deep learning in gravitational-wave astronomy

2 stars 1 forks source link

A bespoke loss function that optimised detection at fixed false alarm #26

Open chrismessenger opened 6 years ago

chrismessenger commented 6 years ago

The current loss function minimises the entropy however we can pick any function we like.

For instance we could only minimise the entropy for the signal class and ignore the noise class.

It might be useful to ask the network to optimise on the detection probability (true alarm) while minimising the abs(false alarm - desired false alarm)? It would then aim to get a certain fraction of noise events wrong while trying it's best to get as many signal right.

Just a thought.

mj-will commented 6 years ago

I've got a basic version working, but there's still a couple of bugs when using it with the theano shared variables. So I'll update when I fix that.

chrismessenger commented 6 years ago

Sounds interesting.