Beers, Floris van (2021) Hedge Backpropagation in Convolutional Neural Networks. Master's Thesis / Essay, Artificial Intelligence.
|
Text
Master_Thesis_Floris_van_Beers.pdf Download (2MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (132kB) |
Abstract
In the field of object recognition much is gained by the use of convolutional neural networks (CNNs). Research into deeper networks has revealed untold successes, as well as unforeseen issues. Ensembles of deep CNNs are being used to further improve performance, while the issues with vanishing gradients have been tackled by the use of deep supervision. In this work a novel architecture is proposed which combines these techniques. Using ResNet34 and DenseNet121 as base variants, a Multiple Heads (MH) adaptation attempts to improve performance and solve issues. Further work on the weights ($\alpha$) in the MH variant leads to use of the Hedge Back Propagation (HBP) algorithm in the HBP and Thaw model variants. Experiments on CIFAR10 and the Naturalis Papilionidae datasets show the use of MH variants improves over base networks in one of the experimental settings. The application of HBP does not further improve the performance of the MH variant, but leads to interesting observations resulting in a multitude of directions for future work.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Jaeger, H. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 29 Nov 2021 12:44 |
| Last Modified: | 29 Nov 2021 12:44 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/26319 |
Actions (login required)
 |
View Item |