Beyond Stationarity: Rethinking Codebook Collapse in Vector Quantization
Hao Lu, Onur C. Koyun, Yongxin Guo, Zhengjie Zhu, Abbas Alili, Metin Nafi Gurcan, [4pt] Wake Forest University School of Medicine, Winston-Salem, NC, USA[4pt] {Hao.Lu, Onur.Koyun, Abbas.Alili}@advocatehealth.org{Yongxin.Guo, Zhengjie.Zhu, Metin.Gurcan}@wfusm.edu · 2026
Read Beyond Stationarity: Rethinking Codebook Collapse in Vector Quantization by Hao Lu, Onur C. Koyun, Yongxin Guo, Zhengjie Zhu, Abbas Alili, Metin Nafi Gurcan, [4pt] Wake Forest University School of Medicine, Winston-Salem, NC, USA[4pt] {Hao.Lu, Onur.Koyun, Abbas.Alili}@advocatehealth.org{Yongxin.Guo, Zhengjie.Zhu, Metin.Gurcan}@wfusm.edu (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-2602-18896.txt.
Description
Vector quantisation powers many generative models, yet its codebooks often collapse. Most code vectors go unused in training. Hao Lu, with five co-authors, offers a new explanation in 2026. The cause is the nonstationary nature of encoder updates. As the encoder drifts, unselected codes receive no updates and slowly die.
Two methods address the problem. Non-Stationary Vector Quantisation propagates encoder drift to non-selected codes through a kernel rule. Transformer-based Vector Quantisation adapts the whole codebook with a lightweight mapping, while preserving convergence to the k-means solution. Tests on CelebA-HQ show near-complete codebook use and better reconstruction than baseline variants. The work gives a principled, scalable basis for future quantisation-based models.
Catalog record
| CREATOR | Hao Lu, Onur C. Koyun, Yongxin Guo, Zhengjie Zhu, Abbas Alili, Metin Nafi Gurcan, [4pt] Wake Forest University School of Medicine, Winston-Salem, NC, USA[4pt] {Hao.Lu, Onur.Koyun, Abbas.Alili}@advocatehealth.org{Yongxin.Guo, Zhengjie.Zhu, Metin.Gurcan}@wfusm.edu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-18896