Discussion of: Learning Classical and machine learning methods for event reconstruction in NeuLAND

David:

• "In turn, FAIRRoot is based on ROOT, an omnipresent software package in nuclear and particle physics"
• Apache Parquet format (preferred for speed and file size)
• 4 different feature sets: One with just 3 cumulative values, one with info for each bar, one that combines the two, and one that identifies hits in pixels (with time and energy being the "colors")
• "All models performed worse when working with the thou- sands of inputs of the Bars dataset than with the three values of the Tri dataset."
• "some models ... perform significantly worse when trained on the number of primary neutrons instead of the (actually detected) number of primary hits."
• Perhaps more time could have been spent on feature engineering?