Deep Belief Network on Context of Handwritten Words

J.J.A. Engelberts (2015) Deep Belief Network on Context of Handwritten Words. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AI_BA_2015_JonneEngelberts.pdf - Published Version Download (519kB) | Preview
	Text Toestemming.pdf - Other Restricted to Backend only Download (47kB)

Abstract

In this paper feature vectors from images of words were used to test whether these feature vectors could be used to cluster semantically similar words in a Deep Belief Network through unsupervised learning. The feature vectors used in this paper came from the handwriting recognition system Monk. First the dimensionality of the feature vectors were reduced with a Restricted Boltzmann Machine. With a reduction of 99% the size of the vectors went from 4356 features to 50 features, without a big loss in Nearest Neighbor finding words with the same label: 75% correct for the 4356 features and 73% correct for the 50 features. To test the ability of the Deep Belief Network clustering semantically similar words a basic artificial language was created. Every sentence in this language had three words per sentence and belonged to one of the two types of context: people or food. The words in the language were first represented by simple Softmax units, which did need a label of the word, but were able to cluster the two contexts perfectly. Secondly the words were represented by a feature vector from an image of that word, this resulted in a pretty decent clustering. Finally the words were represented by ten feature vectors from different images of that word, but this did not results in any clustering in the network. Using feature vectors instead of Softmax units had a slightly worse performance, but when using ten feature vectors per word the network was unable to cluster the two contexts.

Item Type:	Thesis (Bachelor's Thesis)
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	15 Feb 2018 08:06
Last Modified:	15 Feb 2018 08:06
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/13124

Actions (login required)

View Item