{"id":"arxiv/2602.19799","url":"https://arxiv.org/abs/2602.19799","tags":["stat.ML","cs.LG","math.OC"],"score":0,"title":"Path-conditioned training: a principled way to rescale ReLU neural networks","domain":"stat.ML","source":"arxiv","authors":["Arthur Lebeurrier","Titouan Vayer","Rémi Gribonval"],"reviews":0,"subject":"stat.ML","summary":"Despite recent algorithmic advances, we still lack principled ways to leverage the well-documented rescaling symmetries in ReLU neural network parameters. While two properly rescaled weights implement the same function, the training dynamics can be dramatically different. To offer a fresh perspective on exploiting this phenomenon, we build on the recent path-lifting framework, which provides a compact factorization of ReLU networks. We introduce a geometrically motivated criterion to rescale neural network parameters which minimization leads to a conditioning strategy that aligns a kernel in…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-02-23","reactions":{}}