Dice Question Streamline Icon: https://streamlinehq.com

Structure of CPWL parameter variety constraints for ReLU networks

Determine the precise structure of the polynomial equations and inequalities (of degree at most L+1) defining the semi-algebraic feasible set of canonical continuous piecewise linear (CPWL) parameters that correspond to functions representable by a ReLU network of given architecture (n_0,\ldots,n_L), including how the structure depends on the architecture and the chosen representation sizes n and m.

Information Square Streamline Icon: https://streamlinehq.com

Background

The authors discuss mapping ReLU network parameters to a canonical CPWL parameterization capturing gradients and intercepts of linear pieces. Prior work establishes existence of polynomial equations and inequalities of bounded degree describing the representable set.

However, the detailed architecture-dependent form of these constraints is not known; clarifying this would bridge the CPWL parameter perspective with the pattern and output varieties developed in the paper.

References

The structure of the equations and inequalities depending on the ReLU network architecture and the choice of $n$ and $m$ remains an open problem for further study that is related to the pattern varieties described above.

Constraining the outputs of ReLU neural networks (2508.03867 - Alexandr et al., 5 Aug 2025) in Section 3.3 (CPWL parameter variety)