Residual Quantization with Implicit Neural Codebooks

Iris A.M. Huijben, Matthijs Douze, Matthew J. Muckley, Ruud Van Sloun, Jakob Verbeek
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:20682-20699, 2024.

Abstract

Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-huijben24a, title = {Residual Quantization with Implicit Neural Codebooks}, author = {Huijben, Iris A.M. and Douze, Matthijs and Muckley, Matthew J. and Van Sloun, Ruud and Verbeek, Jakob}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {20682--20699}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/huijben24a/huijben24a.pdf}, url = {https://proceedings.mlr.press/v235/huijben24a.html}, abstract = {Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.} }
Endnote
%0 Conference Paper %T Residual Quantization with Implicit Neural Codebooks %A Iris A.M. Huijben %A Matthijs Douze %A Matthew J. Muckley %A Ruud Van Sloun %A Jakob Verbeek %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-huijben24a %I PMLR %P 20682--20699 %U https://proceedings.mlr.press/v235/huijben24a.html %V 235 %X Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.
APA
Huijben, I.A., Douze, M., Muckley, M.J., Van Sloun, R. & Verbeek, J.. (2024). Residual Quantization with Implicit Neural Codebooks. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:20682-20699 Available from https://proceedings.mlr.press/v235/huijben24a.html.

Related Material