{"id":"arxiv/2610.38938","url":"https://arxiv.org/abs/2610.38938","tags":["stat.ML","cs.LG","math.ST"],"score":0,"title":"Memorization Capacity of Invariant Convolutional Neural Nets","domain":"stat.ML","source":"arxiv","authors":["Sjoerd Dirksen","Patrick Finke","Paul Geuchen"],"reviews":0,"subject":"stat.ML","summary":"Neural networks are powerful memorizers which are often able to interpolate their training data even when the labels have been randomized. While there exists an extensive literature on the memorization capacity of fully connected networks, in practice more specific architectures with problem-dependent properties are preferred, such as convolutional networks for image classification. In this paper we consider convolutional networks which are invariant with respect to a given arbitrary finite group of orthogonal transformations, which may include translations, reflections, and/or rotations of…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-10-28","reactions":{}}