Table 1 Results of the transfer learning experiments.

From: Learning to rank Higgs boson candidates

Precise data sets (KS Values)

CNN

CNN + DirectRanker

Retrained CNN + DirectRanker

CMS detector (0.003)

\(2.236 \pm 0.013\) \(Z_0\)

\(2.377 \pm 0.015\) \(Z_0\)

\(2.387 \pm 0.008\) \(Z_0\)

ATLAS-2.1T (0.018)

\(2.288 \pm 0.020\) \(Z_0\)

\(2.348 \pm 0.026\) \(Z_0\)

\(2.357 \pm 0.030\) \(Z_0\)

WW2j/Z2j background (0.002)

\(2.433 \pm 0.023\) \(Z_0\)

\(2.787 \pm 0.033\) \(Z_0\)

\(2.763 \pm 0.025\) \(Z_0\)

  1. The first column of the table shows the generated precise data sets. By calculating the mean of the Kolmogorov–Smirnov32 (KS) test over all 32 features, the similarity of the precise data to imprecisely generated \(t\bar{t}\) and Higgs events is shown in brackets. For the precise data sets, we generated \(t\bar{t}\) and Higgs events with the CMS detector (CMS detector), \(t\bar{t}\) and Higgs events with the ATLAS detector having a 2.1T magnetic field (ATLAS-2.1T) and WW2j/Z2j events with the ATLAS detector. The imprecise generated data uses the normal ATLAS detector. The other two columns show the models we compare to our retrained model. Both models are only trained and cross-validated on the precise data sets. Therefore, we use the CNN and the CNN + DirectRanker model. For our retrained model, we first trained CNN layers on imprecisely generated \(t\bar{t}\) and Higgs data and then we retrained the full CNN + DirectRanker model on a subsample of the precise data sets. The results for this is shown in the last column. For all model results we report the \(Z_0\) metric.
Back to article page