As an alternative to the likelihood method of event classification I've been working on, I have recently built a neural network for the same purpose, namely to separate signal from background. Here are the variables I put into the network. As always, green is atmospheric MC, red is cosmic MC. The distributions before any net cut are normalized to unit area for ease of shape comparison.
| Variable | Distribution Before Net | Distribution After Net |
|---|---|---|
| Number of strips in event | |
|
| Track vertex Y | |
|
| Track vertex y direction cosine | |
|
| Track vertex z direction cosine | |
|
| Track end y direction cosine | |
|
| Track end y direction cosine | |
|
| Track vertex z trace | |
|
| Vertex Charge | |
|
| Event Charge Per Plane | |
|
| Fraction of Event Charge in Track | |
|
The neural network, when trained on a sample of events which pass the usual quality, kinematic, and fiducial cuts, achieves with the quality and fiducial cuts a sig*sig/(sig+back) of 67.3 with 15.5% signal acceptance, 9.3x10^-6 background acceptance, and signal:background of 0.54:1 (cosmic and atmospheric MC samples are properly normalized and no veto shield cuts applied). As usual, cosmic MC is shown in red and atmospheric MC is shown in green. The horizontal axis is the net output, which should be 0 for cosmic and 1 for atmospheric.
The following plot shows the FOM as a function of the cut on neural net output.
| Cosmic MC | Atmos MC | Data | |
| Total | 3.82x10^7 | 1237.85 | 4.11x10^7 |
| Quality | 3.25x10^7 | 619.77 | 3.10x10^7 |
| Kinematic | 1.40x10^7 | 578.15 | 1.27x10^7 |
| Fiducial | 18566.60 | 258.83 | 18391 |
| Neural Net | 355.21 | 191.87 | 728 |