dc.contributor.advisor | Downing, Keith | |
dc.contributor.advisor | Misimi, Ekrem | |
dc.contributor.advisor | Måløy, Håkon | |
dc.contributor.author | Aunrønning, Ola | |
dc.date.accessioned | 2019-09-11T10:56:42Z | |
dc.date.created | 2018-06-11 | |
dc.date.issued | 2018 | |
dc.identifier | ntnudaim:18788 | |
dc.identifier.uri | http://hdl.handle.net/11250/2615884 | |
dc.description.abstract | Deep neural networks are black boxes. While we know how they learn, we still don t have a great understanding of what they learn. This project has a goal of visualizing and understanding what convolutional neural networks learn during training. We aim to develop an approach and methodology for visualization and then apply to a specific dual stream recurrent network (DSRN) used for video action recognition. The DSRN architecture classifies underwater videos of salmon, recognizing whether they are feeding or not. The two streams capture spatial and temporal data separately, and we applied our technique on both streams.
A literature review showed that the visualization techniques used today are primarily focused on convolutional neural networks for image classification. The work in this thesis regarding visualization of architectures for video action recognition is novel. By using image space gradient ascent we were able to isolate what features were able to isolate what features then network is looking for. A generative adversarial network was trained to increase intelligibility of the isolated features. The visualizations also provides some insight as to how we humans can differentiate between feeding and nonfeeding fish. In addition to this, we discovered architectural weaknesses by investigating the distribution of activations. By altering the distribution, we were able to increase the spatial stream accuracy from 72.2% to 79.4% | en |
dc.language | eng | |
dc.publisher | NTNU | |
dc.subject | Datateknologi, Kunstig intelligens | en |
dc.title | Understanding and Visualizing Filters in Deep Convolutional Neural Network Architectures | en |
dc.type | Master thesis | en |
dc.source.pagenumber | 84 | |
dc.contributor.department | Norges teknisk-naturvitenskapelige universitet, Fakultet for informasjonsteknologi og elektroteknikk,Institutt for datateknologi og informatikk | nb_NO |
dc.date.embargoenddate | 10000-01-01 | |