The HEP-ML Decoder: Every Machine-Learning Term, Translated Into Physics You Already Know
If you were trained in particle physics and then handed a machine-learning task, the hardest part is often not the mathematics — it's the vocabulary. ML papers and tutorials are written by and for computer scientists, so they name familiar ideas with unfamiliar words.
The good news: almost every core ML term maps one-to-one onto something you already use every day. Here is the decoder.
The core translation table
| Machine-learning term | What it means in physics |
|---|---|
| Feature | An observable — an input variable like transverse momentum, impact parameter, or an isolation sum |
| Label / target | The truth category from Monte Carlo (signal vs background), i.e. what you truth-match against |
| Model | A parametrised function you fit to data — your χ² fit is already one |
| Training | Fitting the parameters — the minimisation step |
| Loss function | The thing you minimise — the ML cousin of your χ² |
| Overfitting / overtraining | Fitting statistical noise instead of the underlying shape — exactly like adding too many free parameters to a fit |
| Regularisation | Penalising complexity to avoid overfitting — like constraining a fit or adding a prior |
| Generalisation | How well the model does on events it never saw — the whole point of a held-out sample |
| Hyperparameter | A knob you set by hand rather than fit (tree depth, learning rate) — like your binning or fit range |
| Cross-validation | Rotating which events train and which test, so every event is scored blind |
Three translations worth dwelling on
"Overtraining" is overfitting a χ²
Every physicist has seen a fit with too many free parameters snake through the data points and then predict nonsense between them. That is overfitting. A boosted decision tree with too many trees, or trees grown too deep, does the same thing: it memorises the training events — including their statistical fluctuations — and then performs worse on fresh data. The diagnostic is the same intuition, too: compare the model's behaviour on the sample it was fit to versus a sample it never saw. If they diverge, you've overfit.
"ROC curve" is efficiency vs purity, rotated
The Receiver Operating Characteristic curve plots true-positive rate against false-positive rate as you sweep a threshold. Rename the axes and it's an object you already live with: signal efficiency on one axis, background rejection (the complement of contamination) on the other. Sweeping the classifier's output cut is exactly sweeping a cut on a discriminating variable — the ROC curve is just the efficiency–purity trade-off drawn for every possible working point at once. A classifier that improves the ROC is one that gives you more signal efficiency at the same background rejection. That is the whole claim behind using BDTs: Roe et al. (NIM A 543 (2005) 577–584) measured 30–50% efficiency gains at equal background rejection versus rectangular cuts.
"Features" are observables — and feature engineering is what you already do
When you compute an invariant mass from four-vectors, or build an isolation variable, or take a ratio of energies, you are doing feature engineering: constructing inputs that make the signal easier to separate. ML didn't invent this. It just gave it a name and reminded you that the choice of inputs often matters more than the choice of algorithm.
Why this matters
Once the vocabulary stops being a wall, the ML literature opens up. You realise you are not learning a foreign subject — you are learning new names and a few new tools for a job you have been doing since your first fit: separating signal from background, and knowing how much to trust the result.
That is exactly where a structured path helps. If you've been told to build a BDT and you want the whole route — from this vocabulary to a validated, publication-ready classifier — start with the free Module 1 and the BDT Roadmap. Module 1 is Introduction to ML for Physicists, and it spends its first pages exactly here: turning the ML dictionary into physics you already own.
Educational material. It does not replace the internal review and approval procedures of your experiment's collaboration.