Abstract

The objective of neural network quantization is to learn networks while their parameters (and activations) are constrained to take values in a small discrete set, usually binary, which is highly desirable for resource-limitted devices and/or real-time applications. In this project, we develop new algorithms for neural network quantization and make steps towards understanding the generalization and robustness properties of such quantized networks.

Publications

Angle Penalized Neural Network Scaled Quantization.
Luis Guerra, Thalaiyasingam Ajanthan, Bohan Zhuang, Ian Reid, and Tom Drummond.
Under review, March 2020.

Mirror Descent View for Neural Network Quantization.
Thalaiyasingam Ajanthan*, Kartik Gupta*, Philip H. S. Torr, Richard Hartley, and Puneet K. Dokania.
International Conference on Artificial Intelligence and Statistics (AISTATS), April 2021.
[pdf] [supp] [arxiv] [code] [bib]

Improved Gradient based Adversarial Attacks for Quantized Networks.
Kartik Gupta, and Thalaiyasingam Ajanthan.
CVPR Workshop: Adversarial Machine Learning in Computer Vision, June 2020.
[pdf] [arxiv] [bib]

Proximal Mean-field for Neural Network Quantization.
Thalaiyasingam Ajanthan, Puneet K. Dokania, Richard Hartley, and Philip H. S. Torr.
International Conference on Computer Vision (ICCV), October 2019.
[pdf] [supp] [arxiv] [poster] [talk] [code] [bib]