List and prioritise possible applications/use cases
Created by: asogaard
Proposal by @asogaard:
Graph-level classification: Either topological, inspired by Naoko's talk at the fall CM 2021 [1]:
- Skimming (
) - Stopping track ()
- Throughgoing track ()
- Starting track ()
- Cascade ()
- Double-bang cascade ()
- Noise
or by process:
-
(throughgoing or stopping track)
-
(cascade)
-
(starting track)
-
(cascade, double-bang)
- Noise
Specific classification tasks will then be concerned with (a subset of) these categories. Which categories are relevant depend on the detector (e.g. vs. not relevant/possible for IceCube-86) and the reconstruction level (e.g. vs. more relevant than flavour classification at earlier filtering levels).
Graph-level regression:
- Neutrino reconstruction
Node-level classification:
- Hit cleaning
Additional/techniques:
- Training to mitigate systematic uncertainties (including variations in training, or using adversarial training)
- General pre-training on surrogate task (cf. BERT)