PTNQ: Post-Training Non-Linear Quantization

Diogo Venâncio, Nuno P. Lopes

Keywords: quantization

TL;DR: A novel post-training quantization technique that searches for the best function through a pool of methods

Abstract:

Quantization is one of the leading techniques to reduce the memory usage of machine learning models. It works by approximating the weights of a model by some function with a smaller domain (e.g., replace 32-bit floats with 8-bit integers that are coefficients in some function that maps back to 32-bit floats).

Although most quantization methods approximate weights with a linear or affine function, the weights of current machine learning models often exhibit non-linear behavior at the extremities. Moreover, some studies suggest that the extremities are important for the end-to-end accuracy.

In this paper, we introduce PTNQ, a novel post-training quantization technique that approximates weights by searching through a pool of non-linear functions. We show that PTNQ provides significant advantages over affine functions, achieving similar accuracy while requiring 2 to 4 fewer bits per coefficient.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1078