Data Loader

We use a custom data loader see code here to construct batches so that the number of background and signal events are even. Additionally, we enforce the batches to also have the same number of events per Njet bin. Note that no restriction is placed on the number of events from each of the varying signal mass hypotheses as the network already has good performance on the high mass, low statistics samples. Additionally, no requirement is placed on balancing the number of nominal ttbar vs systematically varied ttbar samples.

An example batch is shown below:

• 2048 Background events:

• Njet = 7: 410 events
• Njet = 8: 410 events
• Njet = 9: 409 events
• Njet = 10: 409 events
• Njet = 11: 409 events

• 2048 Signal events:

• Njet = 7: 410 events
• Njet = 8: 410 events
• Njet = 9: 409 events
• Njet = 10: 409 events
• Njet = 11: 409 events

Note that we use a batch size of 4096 for all of our trainings as it has shown the best balance between signal vs. background discrimination and closure performance.