Sparse Federated Echo State Networks for Time-Series Classification via Smooth-L1 Regularization
Lyu, Röder and Schleif · arXiv 2610.21742
The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.
Claim tested
At one fixed threshold, the sparse readouts are 77 to 95% sparse (centralised) and 75 to 95% sparse (federated) and, on five of seven benchmarks, within 0.7 and 1.6 accuracy points of the stronger dense readout.
What we ran and what came back
Seven benchmark datasets, 20 new seeds for the centralised setting and 30 for the federated one, at a single fixed threshold. The run used the first arXiv version and the code as released then. The sparse readouts were at least 70% sparse on all seven datasets in both settings. Accuracy stayed within 1.5 points of the better dense readout on 5 of 7 datasets (centralised) and within 2.5 points on 6 of 7 (federated); these are the same datasets the paper reports as close. Our margins are about twice the 0.7 and 1.6 points in the paper's abstract, and we set them before the run because the paper's own seed-to-seed spread is larger than those figures. Read as bounds that should hold on new seeds, with a seed-paired test, 0.7 and 1.6 are not met: 1 of 7 datasets is within 0.7 (centralised) and 4 of 7 within 1.6 (federated). The current arXiv version (v3, 4 November 2026) states that these figures are differences of ten-seed means, of the same order as the seed-to-seed variation. On that definition our means give 1.33 for the federated setting (within 1.6) and 1.00 for the centralised one (against 0.7). The pattern of sparsity and near-dense accuracy reproduces; the two figures are best read as descriptions of the authors' runs, not as bounds.
What this result supports
At a single fixed threshold the readouts are at least 70% sparse on all seven benchmarks in both settings, and accuracy stays within 1.5 points (centralised, 5 of 7 datasets) and 2.5 points (federated, 6 of 7) of the better dense readout.
What it does not support
That 0.7 and 1.6 points hold as bounds on new seeds; anything about other datasets, thresholds or reservoir sizes.
Record
Protocol, frozen by git commit before the run and deposited afterwards (the archive was unreachable at the time): 10.5281/zenodo.23452631
The authors were sent this result. They have not reviewed this listing.
How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.