Calorimeter Machine Learning

Name: Calorimeter Machine Learning
Start: 2018-04-20T15:00:00+02:00
End: 2018-04-20T17:40:00+02:00
Location: No location set

Friday 20 Apr 2018, 15:00 → 17:40 Europe/Zurich

Hide

Gulrukh

Junze

Ran GoogLeNet in TriForce and got 95% accuracy with batch normalization (previously got 99% with feature scaling)
Action item: repeat this check of DNN and GoogLeNet for electron vs. charged pion
Action item: try to reduce overtraining:
- Run with more events
- Enable dropout (or increase dropout rate)
- Try L1 or L2 regularization in training (Maurizio to send paper on reducing overtraining, see also https://discuss.pytorch.org/t/simple-l2-regularization/139)
Action item: for new fixed-angle samples, check dependence of performance, speed, memory usage, etc, on window size

Matt

There are minutes attached to this event. Show them.

Choose timezone