Decomposition Polyhedra of Piecewise Linear Functions
Researchers have developed a new framework for decomposing piecewise linear functions into the difference of two convex functions. This work addresses a challenge in optimization and neural network theory, where finding decompositions with minimal linear pieces is crucial. The study disproves a prior approach and introduces a method that fixes the polyhedral complex underlying the function's nonlinearity, proving that decompositions form a polyhedron and minimal solutions correspond to its vertices. This framework has implications for submodular functions and improves neural network constructions for convex and nonconvex piecewise linear functions. AI
IMPACT Provides a new theoretical tool for constructing and analyzing neural networks, potentially improving their efficiency and capabilities.