Lohuizen, Quintin van (2021) Training Deep Neural Networks with Soft Loss for Strong Gravitational Lens Detection. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI_2021_LohuizenQv.pdf Download (6MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (124kB) |
Abstract
This thesis explores a binary classification problem using convolutional neural networks (CNNs) applied to gravitational lensing. We try to find strong gravitational lenses in the Kilo-Degree Survey (KiDS) dataset, where we lack large labelled training sets and have noisy data. It can lack sufficient resolution and is severely imbalanced. Currently, astronomers spend a significant amount oftime manually labelling images with a large portion of false positives. CNNs can reduce the strain on astronomers and speed up the process by rejecting obvious negative cases. This work investigates various loss functions and their influence on the trade-off between precision and recall. We consider the following loss functions: binary cross-entropy, F1-, F1 double- and Fβ soft loss. Our results show that binary cross-entropy reaches the highest accuracy, while Fβ reaches the highest precision and F1 soft loss the highest recall. The loss function choice can influence how much time an astronomer would need manually labelling false positives. We used simulated lensing features to address a lack of real-world labelled data and merged them with real luminous red galaxies (LRGs). Merging lens and source enables us to understand the relation between misclassification and the lensing parameters’ values. Our findings suggest that low brightness and size cause most false-negative cases and not other galaxies in the image’s background.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Wiering, M.A. and Koopmans, L.V.E. and Verdoes Kleijn, G.A. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 19 Apr 2021 09:09 |
| Last Modified: | 19 Apr 2021 09:09 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/24273 |
Actions (login required)
 |
View Item |