Present: Aris, Jim, Mayur, Stephan, Steve, Josh, Paolo
Jean-Roch will present HEP.TrkX at ATLAS ML workshop Jun 6-9.
FNAL
Aris started playing with latest Dustin LSTM network similar to what JR did track by track instead than globally for an event
Giuseppe started working on loss function for training
Caltech
Josh’s generative network talk at SLAC last week
https://indico.cern.ch/event/568875/contributions/2397925/attachments/1459058/2253175/mcgbr-May12-2017.pdf
LBL
Steve’s Continuous space hit classifier
Fully connected NN, works qualitatively well picking one track (10 hits) out of 50 hits (5 tracks). Slow to converge (50M samples to train). LSTM (training as we speak) is faster but still not fully converged before the meeting.
CNN trains even better, and performs well despite being an atypical application for a CNN.
There are minutes attached to this event.
Show them.