Abstract
dc:descriptionNeural networks are known for their large number of parameters and requiring a lot of computing resources for complex classification tasks. They often are overparametrized, some neurons are not used or are not contributing much to the result. If low importance neurons could be ranked by how much they contribute, those useless neurons could be pruned. Pruning would lead to faster and smaller neural networks. In this paper, investigating how much networks can be pruned without losing accuracy or losing a small part of it and optimizing neural network resource usage. This paper demonstrates, that there is a point, where accuracy drops significantly. Also, shown, that for pruned structure old weights are not that important, as it was expected.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Macikas, Tomas,
- Contributors dc:contributor
-
- Kybartas, Rimantas
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:ELABAETD81805933&prefLang=en_US