Institutional Repository of Vilnius University
Gilaus mokymo dirbtinių neuroninių tinklų taikymas besikeičiančio uždavinio sprendimui /
Abstract
dc:descriptionNowadays one of the main challenge to neural networks — ensure that network performs in generalised manner [GBC16]. Neural network should work efficiently even with data which features it had not learned explicitly. However, it is still unclear, how different circumstances, which appear at network learning and testing process, affect main task completion efficiency and accuracy. This research investigate how new type of Deep Learning Neural Network — “Capsule Neural Network” (CapsNet) can adapt to task operating on objects, whose features could potentially change (In this research changing task was modeled as animal species change). Alongside comparison was made with standard “Convolutional Neural Network” (CNN). It is claimed that CapsNet can achieve better performance on classification and segmentation tasks then traditional CNN [HSF18; SFH17]. With this research it was displayed that capsule neural networks can adapt more easily towards changing tasks when appropriate data augmentation techniques are performed. Additionally, for CNN roll data augmentation help adapt to changing task, while for CapsNet — Gaussian noise, rotation and roll transformations where quite successful. And finally, greater performance for CapsNet was achieved using transformations which mostly result in object features coordinates change (rotation and roll), rather than with transformations which distort actual objects features (blur, motion blur), however exception was observed with Gaussian noise transformation result.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Blaževičius, Marius,
- Contributors dc:contributor
-
- Valaitis, Vytautas
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD29548589&prefLang=en_US