Milanese, Samuele (2022) Transfer Learning with Uncertain Features. Bachelor's Thesis, Artificial Intelligence.
|
Text
report final.pdf Download (3MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (125kB) |
Abstract
As Deep Learning is used in an increasing number of (sensitive) scientific fields and real-world applications, Transfer Learning is being used in those areas where acquiring data is expensive and/or difficult. Although Transfer Learning is a useful technique to counter scarcity of data, it carries DL problems such as over- or under-confidence and uncalibrated predictions in general. In standard settings, Uncertainty Quantification methods can be used to achieve safe and reliable models with calibrated confidence scores (Leibig et al., 2017). This research aims at using different UQ methods to extract uncertain features (with mean and variance) instead of point features (only mean) in order to carry the advantages of UQ to Transfer Learning. This is explored by building and comparing feature extractors in three different setups: different UQ methods applied to different architectures; models with only some of the layers implementing uncertainty quantification; uncertain features are sampled to generate new data points. The quality of the features is evaluated by feeding them to an SVM (Ho & Kim, 2021). The results of the experiments did not show any improvement in performance when using uncertain rather than point features, neither in terms of accuracy nor expected calibration error. However, generating new data through sampling uncertain features could be suggested as a valid Data Augmentation technique.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Valdenegro Toro, M.A. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 16 Aug 2022 09:07 |
| Last Modified: | 16 Aug 2022 09:07 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/28363 |
Actions (login required)
 |
View Item |