Soft-Radial Projection for Constrained End-to-End Learning
Philipp J. Schneider Daniel Kuhn Risk Analytics and Optimization Chair, EPFL {philipp.schneider, daniel.kuhn}@epfl.ch · 2026
Read Soft-Radial Projection for Constrained End-to-End Learning by Philipp J. Schneider Daniel Kuhn Risk Analytics and Optimization Chair, EPFL {philipp.schneider, daniel.kuhn}@epfl.ch (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03461.txt.
Description
Safety-critical learning systems must respect hard constraints, not just prefer them. Philipp J. Schneider and Daniel Kuhn address a hidden flaw in the usual fix, projecting predictions back onto the allowed set, in their 2026 paper. Standard projection flattens outside points onto a boundary surface, which kills the gradients that training depends on.
Their Soft-Radial Projection layer instead maps points radially inward, landing well inside the feasible set. Every output stays strictly feasible, while the Jacobian keeps full rank almost everywhere, so optimization no longer stalls. The authors prove the architecture keeps universal approximation power. Experiments show better convergence and solution quality than optimization-based and projection-based baselines.
Catalog record
| CREATOR | Philipp J. Schneider Daniel Kuhn Risk Analytics and Optimization Chair, EPFL {philipp.schneider, daniel.kuhn}@epfl.ch |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 97 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-03461