Institutional Repository of Vilnius University
Konvoliucinių neuroninių tinklų rezultatų analizė remiantis vizualizacijos metodais /
Abstract
dc:descriptionArtificial neural networks have become one of the most advanced and widely used technologies in many fields. However, their black box architecture poses challenges for their application in real‑world domains where solution security and transparency are crucial. Neural network visualizations address this issue by ensuring accountability, fairness, compliance with laws and regulations, and security when deploying artificial intelligence systems. This study involves training three YOLOv8 classification models with different performance metrics and creating and comparing their output class activation maps. Methods such as pixel count, root mean square deviation, structural similarity index, and histogram intersection were used to obtain metrics used to search for correlations between class activation maps and specific result accuracy. Comparing EigenCAM maps applied to all three models using the described image comparison methods did not reveal any strong dependencies between the methods used and model confidence scores. This refutes the hypothesis that convolutional neural network visualizations can be used to determine the accuracy of a specific YOLOv8 convolutional neural network result for the FF++ dataset. Exploration of the correlation between specific result accuracy and quantile activations in EigenCAM maps revealed that in the dataset used, the 60th quantile is the most significant when analyzing the areas of interest in class activation maps.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Rėkus, Danielius,
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:ELABAETD210647901&prefLang=en_US