| Title | Investigation of generalization properties of convolutional neural networks / |
|---|---|
| Translation of Title | Konvoliucinių neuroninių tinklų generalizavimo savybių tyrimas. |
| Authors | Ščigla, Edvinas |
| Full Text |
|
| Pages | 55 |
| Keywords [eng] | generalizavimas, konvoliuciniai neuroniniai tinklai, neuroninio tinklo architektūra, viso vaizdo duomenų augmentacija, vaizdo skaldymas ir duomenų augmentacija kiekvienai išskaldytai daliai, sujungti tinklai. generalization, convolutional neural networks, neural network architecture, data augmentation on image level, data augmentation on non-image level, ensembled networks. |
| Abstract [eng] | In this master’s thesis the analysis of the current state of generalization properties on convolutional neural networks was done, investigated the impact of plain convolutional neural network architecture (ResNet and DenseNet with different number of layers) on convolutional neural network training and generalization and compared with ensembled network models. Investigated the impact of data augmentation on image level (CIFAR-10 and CIFAR-100 datasets), implemented data augmentation on non-image level and tested what impact it has to generalization compared with image level results. Conducted experiments allowed us to draw a conclusion that: data augmentation on non-image level brings better generalization compared to traditional data augmentation on image level. Both ensembled network models (ResNet and DenseNet) showed better results on generalization compared with plain convolutional neural network architectures, because individually each convolutional neural network has its own weaknesses, but together when they aggregated to generate single output – they are showing slightly better generalization results. |
| Dissertation Institution | Vilniaus universitetas. |
| Type | Master thesis |
| Language | English |
| Publication date | 2023 |