Institutional Repository of Vilnius University
Giliųjų neuroninių tinklų atkartojamumas /
Abstract
dc:descriptionThis bachelor’s thesis investigates the effect of random initialisation on deep neural networks and the stability of results. It also investigates noisy data impact on models accuracy. This thesis is es pecially relevant nowadays, amid the AI boom. Corporate investment in AI in 2024 was 252.33 billion dollars, while in 2025 this number rose to 581.69 billion dollars, according to the Stanford University annual artificial intelligence index report. This rapid growth increases the need to examine the issues related to this technology’s reliability, stability and reproducibility. It is known that initialisation af fects the model’s results but this connection is not sufficiently studied. That became one of the main goals of this thesis. To achieve the goals of this thesis we formulated 5 main tasks: 1. To review the literature related to the reproducibility and stability of deep neural networks. 2. Create and train a large number of deep neural networks, initializing them at different seed intervals. 3. Test whether there are significant differences in accuracy results between different seed inter val groups. 4. Test whether there are significant differences in confusion matrices between different seed interval groups. 5. Investigate the impact of noisy data on the accuracy of models. Experiment started with the creation and training of 1000 convolutional neural networks. They were trained to classify images from the FashionMNIST dataset. All of the nets had the same archi tecture, the only difference between them was the seed, which affected values initialisation. Ne tworks were divided into 10 groups based on exponentially increasing seed intervals. Groups were formed to test the seed’s impact on the model’s results later. Each network was tested and its ac curacy and confusion matrices were saved. • The ANOVA statistical method was used to compare the accuracy of networks between groups. • The Welch ttest was used to compare the accuracy of networks within groups. • The PERMANOVA statistical method was used to compare the confusion matrices both be tween groups and within groups. None of the tests found any significant difference between groups or within groups, indicating that these types of networks are reasonably stable with respect to their initialisation. Second part of this experiment was made to investigate artificial noise impact on the accuracy of networks. 280 additional convolutional nets were created but this time they were being trained on noisy data. It was chosen to use horizontal blur as the artificial noise. This type of blur could appear naturally in the photos if an object or the camera was moving at the time of making a photo. Horizontal blur equalised each pixel’s value with its neighbors on the horizontal line. 20 networks were created and trained for each of the blur sizes, which were in the range of 2 to 15 pixels. The bigger the blur size was, the more neighbors affected each pixel. The sample size for this experiment was found using power analysis method. Welch ttest was used to compare accuracy results between original models and models trained on blurry data. It showed that even the smallest blur size affected the net so much, that accuracy results were different by a statistically significant margin. 1% accuracy degradation was achieved at a blur size of 5 pixels. Increasing blur size caused accuracy results to fall faster, especially after blur size of 10 pixels, which suggests that nets rely more on bigger structural features of objects than on subtle details. This also shows that results rely more on the quality of data than on initialisation of the network.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Survilo, Martin,
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:ELABAETD308336906&prefLang=en_US