Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Occlusion method to obtain saliency maps for CNN

Eriksson, Thomas LU (2020) FYTK02 20201
Computational Biology and Biological Physics - Has been reorganised
Abstract
This Bachelor project will study convolutional neural networks created for image classification. Furthermore, it will specifically use an explanatory model for how the network decided a certain classification output. This is to increase the interpretability of the network. However, the completeness of the explanatory model needs to be high for it to be useful. A saliency map of how valuable each image pixel is for the classification will be created, by occluding parts of the image. The MNIST dataset was used, which contains handwritten digits. The main points of research were to study ways to occlude or filter parts of the image. Among the researched topics were the size of the filter, the number of filters and how the filtered pixels... (More)
This Bachelor project will study convolutional neural networks created for image classification. Furthermore, it will specifically use an explanatory model for how the network decided a certain classification output. This is to increase the interpretability of the network. However, the completeness of the explanatory model needs to be high for it to be useful. A saliency map of how valuable each image pixel is for the classification will be created, by occluding parts of the image. The MNIST dataset was used, which contains handwritten digits. The main points of research were to study ways to occlude or filter parts of the image. Among the researched topics were the size of the filter, the number of filters and how the filtered pixels should be alternated. The occlusion method to obtain saliency maps was compared with rivalling methods, such as deepLIFT. The conclusion was that the large amount of computational power needed limits the use of occlusion based methods, but the high completeness makes it useful for niche purposes. (Less)
Popular Abstract
It is today widely accepted that automatization is on its way, with the consequence of a majority of jobs being done by artificial intelligence (AI). One of the leading AI methods is artificial neural networks (ANN). The ANN is a unique method, since it is designed to replicate the human brain. The ANN approach is widely used for voice and text recognition, visual classification, medical diagnostics, and many other applications. The quality of these are essential for the automatization to go smoothly. This project sets out to visualize how the visual classifying ANNs work, which is done with a slightly changing the input and observing the change in output. Automatization will take the jobs from the transportation industry with self-driving... (More)
It is today widely accepted that automatization is on its way, with the consequence of a majority of jobs being done by artificial intelligence (AI). One of the leading AI methods is artificial neural networks (ANN). The ANN is a unique method, since it is designed to replicate the human brain. The ANN approach is widely used for voice and text recognition, visual classification, medical diagnostics, and many other applications. The quality of these are essential for the automatization to go smoothly. This project sets out to visualize how the visual classifying ANNs work, which is done with a slightly changing the input and observing the change in output. Automatization will take the jobs from the transportation industry with self-driving cars. It is then important that the classification these machines does of its environment is done in a correct way, otherwise there could be lethal consequences. Therefore, tools to investigate how the classification works are essential. While this is a simple perturbation-based tool, it could be useful to get a quick look at how the ANN works. (Less)
Please use this url to cite or link to this publication:
author
Eriksson, Thomas LU
supervisor
organization
course
FYTK02 20201
year
type
M2 - Bachelor Degree
subject
keywords
Saliency map, explanatory model, process focussed XAI, MNIST, image classifying CNN.
language
English
id
9023612
date added to LUP
2020-07-14 10:39:59
date last changed
2020-07-14 10:39:59
@misc{9023612,
  abstract     = {{This Bachelor project will study convolutional neural networks created for image classification. Furthermore, it will specifically use an explanatory model for how the network decided a certain classification output. This is to increase the interpretability of the network. However, the completeness of the explanatory model needs to be high for it to be useful. A saliency map of how valuable each image pixel is for the classification will be created, by occluding parts of the image. The MNIST dataset was used, which contains handwritten digits. The main points of research were to study ways to occlude or filter parts of the image. Among the researched topics were the size of the filter, the number of filters and how the filtered pixels should be alternated. The occlusion method to obtain saliency maps was compared with rivalling methods, such as deepLIFT. The conclusion was that the large amount of computational power needed limits the use of occlusion based methods, but the high completeness makes it useful for niche purposes.}},
  author       = {{Eriksson, Thomas}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Occlusion method to obtain saliency maps for CNN}},
  year         = {{2020}},
}