Lavanya Nemani       11:31 AM
@Jeroen, how is the performance measured on unlabelled light curves?

This question has been answered live


Anonymous Attendee       11:33 AM
Just before calculating the entropy, I you mentioned you average things in batches of 10 points. Is 10 just the number that works, or does it come from theory, or did you do cross validation? Thank you

Jeroen Audenaert       11:38 AM
The scaling factor of 10 comes from a combination of theory and practice. In order to obtain consistent entropy values it is recommended to always have a minimum number of datapoints at each scale. So the value of 10 is a good trade-off between this minimum number and still seeing enough detail. We actually also tested with a maximum value of 20, and this was still working nicely as well.


Larry Fitzpatrick       11:36 AM
Question for @Jeroen: Amazing talk!! For TESS, would it be possible to simulate a training data set? from Nikki Arendse. It came thorugh on Slido. Best, Larry

Jeroen Audenaert       11:43 AM
Thanks! Simulating a TESS dataset is definitely a possibility as well. Except for the constant class, we specifically used real data at the moment because it then has the true noise and instrumental properties, whereas simulated data might not be a perfect representation. But maybe combining the iterative approach with simulated data could work nicely as well actually!