All Images

The Big Picture

Figure 1

Example confusion matrix for a 3-class problem.
Example confusion matrix for a 3-class problem.

Figure 2

Example ROC curve.
Example ROC curve.

Working in Google Colab

The Complete Code

Figure 1

Confusion matrix from a full 10-epoch training run of MiniParT.
Confusion matrix from a full 10-epoch training run of MiniParT.

Figure 2

ROC curves from the same full training run.
ROC curves from the same full training run.

What Is a Jet?

Finding the Truth Labels

Figure 1

Truth quarks (filled dots) are matched to nearby reconstructed jets (green triangles) within a ΔR window; jets outside that window (grey triangles) are unmatched and excluded from training.
Truth quarks (filled dots) are matched to nearby reconstructed jets (green triangles) within a ΔR window; jets outside that window (grey triangles) are unmatched and excluded from training.

Preparing the Data

Building MiniParT

Training the Model

Evaluating the Model

Figure 1

Confusion matrix from a full 10-epoch training run of MiniParT. QCD is separated cleanly from the two signal classes (only 295 and 304 events leak into the QCD column), while Hbb and Hcc are confused with each other far more often (2382 and 2865 events respectively). This is the Hbb-vs-Hcc difficulty from The Big Picture episode showing up directly in real results.
Confusion matrix from a full 10-epoch training run of MiniParT. QCD is separated cleanly from the two signal classes (only 295 and 304 events leak into the QCD column), while Hbb and Hcc are confused with each other far more often (2382 and 2865 events respectively). This is the Hbb-vs-Hcc difficulty from The Big Picture episode showing up directly in real results.

Figure 2

ROC curves from the same full training run. QCD vs. Rest reaches an AUC of 0.984, close to a perfect classifier, while Hbb vs. Rest (0.835) and Hcc vs. Rest (0.833) are close to each other and noticeably lower. The two signal curves nearly overlap across the whole plot, another way of seeing that the model finds Hbb and Hcc harder to separate from each other than either is to separate from QCD.
ROC curves from the same full training run. QCD vs. Rest reaches an AUC of 0.984, close to a perfect classifier, while Hbb vs. Rest (0.835) and Hcc vs. Rest (0.833) are close to each other and noticeably lower. The two signal curves nearly overlap across the whole plot, another way of seeing that the model finds Hbb and Hcc harder to separate from each other than either is to separate from QCD.