Increasing the storage capacity of Recursive Auto-associative Memory by segmenting data
2005 (English)Licentiate thesis, monograph (Other academic)
Recursive Auto-associative Memory (RAAM) has been proposed as a connectionist solution for all structured representations. However, it fails to scale up to representing long sequences of data. In order to overcome this, a number of different architectures and representations have been proposed. It is here argued that part of the problem of, and solution to, storing long sequences in RAAM is data representation. It is proposed that by dividing the sequences to be stored into smaller segments that are individually compressed the problem of storing long sequences is reduced.
This licentiate thesis investigates which different strategies there are for segmenting the sequences. Several segmentation strategies are identified and organized into a taxonomy. Also, a number of experiments based on the identified segmentation strategies that aims to clarify if, and how, segmentation affect the storage capacity of RAAM are performed.
The results of the experiments show that the probability that a sequence of a specific length stored in RAAM can be correctly recalled is increased by up to 30% when dividing the sequence into segments. The performance increase is explained by that segmentation reduces the depth at which a symbol is encoded in RAAM which reduces a cumulative error effect during decoding of the symbols.
Place, publisher, year, edition, pages
Linköping: Linköpings universitet , 2005. , 118 p.
Linköping Studies in Science and Technology. Thesis, ISSN 0280-7971 ; 1191
IdentifiersURN: urn:nbn:se:liu:diva-33294Local ID: 19296ISBN: 91-85457-21-3OAI: oai:DiVA.org:liu-33294DiVA: diva2:254117
2005-11-30, G110, Hus G, Högskolan i Skövde, Skövde, 10:15 (Swedish)