Mapping machine-learned physics into a human-readable space

We present a technique for translating black-box machine-learned classifier operating on high-dimensional input space into small set of human-interpretable observables that can be combined to make the same classification decisions. iteratively select these from large high-level discriminants by finding those with highest decision similarity relative black box, quantified via metric we introduce evaluates ordering pairs inputs. Successive iterations focus only subset are misordered current observables. This method enables simplification machine-learning strategy, interpretation results in terms well-understood physical concepts, validation model, and potential new insights nature problem itself. As demonstration, apply our approach benchmark task jet collider physics, where convolutional neural network acting calorimeter images outperforms six well-known substructure Our maps called energy flow polynomials, it closes performance gap identifying class an interesting has been previously overlooked literature.