The Anchor Accumulation Problem: Bounded Anchor Sets Recover Plasticity in Usage-Localized Elastic Weight Consolidation
Version 3 is the Pattern Recognition submission package. Relative to v2 it updates the manuscript (34 pp.) and the accompanying reproducibility package in five ways. (1) Statistics: every headline comparison now reports the two-sided Wilcoxon signed-rank p-value alongside the paired t statistic, plus the Shapiro-Wilk p-value on the paired differences; all four primary comparisons reach Wilcoxon p = 0.002 (the minimum attainable at n = 10), and one cell whose paired differences violate normality (Shapiro-Wilk p = 0.024) is reported explicitly. New script audit/audit_nonparametric.py recomputes these tests from the raw records. (2) References: five works added (four recent Pattern Recognition continual-learning papers plus DER++), all cited; the bibliography now numbers 34 entries, each cited in the text. (3) The manuscript cites the Zenodo concept DOIs, which always resolve to the latest version of each record. (4) Reproducibility: a one-command verification driver (audit/verify_all.py, 8 checks, all passing), a LaTeX pre-flight checker, an academic-English lint, a reproducibility container (Dockerfile, requirements.txt), and audit/REPRODUCE.md; the verification scripts have been run unchanged in an independent Python 3.12 environment. (5) Data: the Split-CIFAR-100 policy-grid record shards (s2r6_b_t10/t20, s2r6_a5) are now included in results/, completing the per-seed records for every number quoted in the CIFAR table. Files: the 34-page manuscript (PDF + full LaTeX source), the submission package (highlights, cover letter, declarations), the graphical abstract, the experiment code, 35 raw per-seed JSON record files, and the audit/ directory now containing twenty verification and figure scripts. Companion repository: github.com/zhouranxi/delp-anchor-budget (to be synced with this version). Related records: precursor DeLP audit doi.org/10.5281/zenodo.23572226 ; S1 segment-specialization audit doi.org/10.5281/zenodo.23637972 ; dataset and code release doi.org/10.5281/zenodo.23639561 (all concept DOIs). --- ## 核对(Publish 前看一眼) - [ ] 文件名仍是 `S2_zenodo_package.zip`(保持一致,别改名) - [ ] Version = 3 - [ ] License 仍为 CC-BY-4.0 - [ ] Resource type 仍为 Preprint - [ ] concept DOI 23668522 出现在页面顶部(别动它,那是系统管的)
Authors
- Ranxi Zhou
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-10-27
- DOI
- https://doi.org/10.5281/zenodo.23668522
- Citations
- 13
- Primary Topic
- Domain Adaptation and Few-Shot Learning
- Type
- preprint