Javascript must be enabled for the correct page display

Scale Selection in Convolutional Neural Networks with Dimensional Min-pooling and Scaling Filters

Vienken, G (2016) Scale Selection in Convolutional Neural Networks with Dimensional Min-pooling and Scaling Filters. Master's Thesis / Essay, Artificial Intelligence.

[img]
Preview
Text
Master_AI_2016_GereonVienken.pdf.pdf - Published Version

Download (4MB) | Preview
[img] Text
Toestemming.pdf - Other
Restricted to Backend only

Download (158kB)

Abstract

Convolutional Neural Networks (CNN) are the most efficient method for image classification and recognition since the year 2012. In the last years they have been applied to more and more complex datasets and continue to hold the records in classification rates for most of them. But in order to perform well on bigger or more complex datasets, CNNs need to have more layers added to them. This makes them more computationally expensive and less biologically plausible. In order to counteract this trend, we propose a new type of convolution layer that makes use of multi-scale convolution filters, which are united with a MinMax function, in order to pass on the results of the most informative filter that is still noise resistant. In order to test this approach, we compare its results to a regular AlexNet implementation, as well as a multi-scale filter approach that uses Max-pooling. Our results show that MinMax can improve the performance of the standard network, even though the final results were not significant with ten iterations, because of overfitting in six of the networks. If we take this overfitting in to consideration and only look at the epochs before it starts, MinMax performed significantly better and improved the network performance by 0.71% on Cifar-10 and 4.9% on Places. This is in line with our expectations that MinMax has a higher impact on complex datasets with a lot of visual information regarding scale differences. These results can be seen as a first step and justify further research into this area, especially to test the impact of this approach when applied to deeper levels of the network.

Item Type: Thesis (Master's Thesis / Essay)
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 15 Feb 2018 08:25
Last Modified: 15 Feb 2018 08:25
URI: http://fse.studenttheses.ub.rug.nl/id/eprint/14651

Actions (login required)

View Item View Item