Please use this identifier to cite or link to this item:
|Title:||Elementwise overparameterisation for single and multi-task learning||Authors:||Ribli, Vincent||Keywords:||Engineering::Computer science and engineering||Issue Date:||2022||Publisher:||Nanyang Technological University||Source:||Ribli, V. (2022). Elementwise overparameterisation for single and multi-task learning. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/156774||Abstract:||In the field of autonomous vehicles, computer vision is used to solve multiple tasks such as semantic segmentation and object tracking. This can be challenging as the tasks need to be done at a high performance within a given latency threshold. Furthermore, multiple tasks need to be solved simultaneously within the given constraints. To solve such an issue, methods such as overparameterisation, along with multi-task learning, have been proposed. This paper proposes a novel overparameterisation technique, along with a few training tricks, which achieves empirically superior performance compared to existing approaches. These ideas are firstly tested on the CIFAR-100 dataset, which is a single-task problem performing image classification. The ideas are further tested on a multi-task setting using the NYUv2 dataset, performing semantic segmentation, depth estimation and surface normals estimation simultaneously. The results of experimentation have shown promising results through the novel overparameterisation approach, and it is hoped that this overparameterisation technique can generalise well to other architectures and datasets as a simple, yet effective approach to improve performance of deep learning models.||URI:||https://hdl.handle.net/10356/156774||Fulltext Permission:||embargo_restricted_20221021||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Student Reports (FYP/IA/PA/PI)|
Files in This Item:
|FYP Report_Vincent Ribli.pdf|
|6 MB||Adobe PDF||Under embargo until Oct 21, 2022|
Updated on Jul 2, 2022
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.