A piecewise-linear neural network \(f\) is a [[ Neural Network ]] for which a bounded input domain \(\mathcal{C}\) can be decomposed into a set of polyhedra \(\mathcal{C}_i\) such that \(\mathcal{C} = \bigcup_i \mathcal{C}_i\) and where the restriction of \(f\) to \(\mathcal{C}_i\) is a linear function for each \(i\).

This type of network represents the majority of networks used in practice.

Sources

• [[ Branch and Bound for Piecewise Linear Neural Network Verification ]]