On Quantizing Implicit Neural Representations

TL;DR

This paper explores advanced quantization techniques for implicit neural representations, showing that non-uniform and clustered quantization significantly improve reconstruction quality and enable highly compressed neural network models.

Contribution

It introduces a novel clustered quantization method for neural weights and demonstrates how it enhances reconstruction and enables ultra-compressed neural representations.

Findings

Clustered quantization improves reconstruction quality.

Binary neural networks can reconstruct signals with high compression.

NeRF can be compressed to less than 16kb with minimal performance loss.

Abstract

The role of quantization within implicit/coordinate neural networks is still not fully understood. We note that using a canonical fixed quantization scheme during training produces poor performance at low-rates due to the network weight distributions changing over the course of training. In this work, we show that a non-uniform quantization of neural weights can lead to significant improvements. Specifically, we demonstrate that a clustered quantization enables improved reconstruction. Finally, by characterising a trade-off between quantization and network capacity, we demonstrate that it is possible (while memory inefficient) to reconstruct signals using binary neural networks. We demonstrate our findings experimentally on 2D image reconstruction and 3D radiance fields; and show that simple quantization methods and architecture search can achieve compression of NeRF to less than 16kb with minimal loss in performance (323x smaller than the original NeRF).