Supervisor

Suitable for

Abstract

Neural fields [1] have emerged as a promising method for representing 3D data, utilising Multi-Layer Perceptrons (MLPs) to predict the properties of a field at every point in space. Despite their potential, a significant drawback is the lengthy training process, often due to the large size of the MLPs. This project aims to explore the application of low-rank adaptation (LoRA) [2] in enhancing the training efficiency of neural fields. An essential part of this investigation will involve benchmarking the performance of low-rank training against full training across various types of fields, such as Signed Distance Functions (SDFs) or radiance fields, to evaluate the effectiveness of LoRA in this context.

[1] https://neuralfields.cs.brown.edu/siggraph23.html

[2] https://arxiv.org/pdf/2106.09685.pdf

Pre-requisites: Suitable for those who have taken a course in machine learning or computer graphics