Explorations of the mean field theory learning algorithm
(1989) In Neural Networks 2(6). p.475-494- Abstract
The mean field theory (MFT) learning algorithm is elaborated and explored with respect to a variety of tasks. MFT is benchmarked against the back-propagation learning algorithm (BP) on two different feature recognition problems: two-dimensional mirror symmetry and multidimensional statistical pattern classification. We find that while the two algorithms are very similar with respect to generalization properties, MFT normally requires a substantially smaller number of training epochs than BP. Since the MFT model is bidirectional, rather than feed-forward, its use can be extended naturally from purely functional mappings to a content addressable memory. A network with N visible and N hidden units can store up to approximately 4N patterns... (More)
The mean field theory (MFT) learning algorithm is elaborated and explored with respect to a variety of tasks. MFT is benchmarked against the back-propagation learning algorithm (BP) on two different feature recognition problems: two-dimensional mirror symmetry and multidimensional statistical pattern classification. We find that while the two algorithms are very similar with respect to generalization properties, MFT normally requires a substantially smaller number of training epochs than BP. Since the MFT model is bidirectional, rather than feed-forward, its use can be extended naturally from purely functional mappings to a content addressable memory. A network with N visible and N hidden units can store up to approximately 4N patterns with good content-addressability. We stress an implementational advantage for MFT: it is natural for VLSI circuitry.
(Less)
- author
- Peterson, Carsten LU and Hartman, Eric
- publishing date
- 1989
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Bidirectional, Content addressable memory, Generalization, Learning algorithm, Mean field theory, Neural network
- in
- Neural Networks
- volume
- 2
- issue
- 6
- pages
- 20 pages
- publisher
- Elsevier
- external identifiers
-
- scopus:0024901271
- ISSN
- 0893-6080
- DOI
- 10.1016/0893-6080(89)90045-2
- language
- English
- LU publication?
- no
- id
- 4dfa588d-1062-4561-a708-17992994cd24
- date added to LUP
- 2019-05-15 07:57:28
- date last changed
- 2021-01-03 11:09:55
@article{4dfa588d-1062-4561-a708-17992994cd24, abstract = {{<p>The mean field theory (MFT) learning algorithm is elaborated and explored with respect to a variety of tasks. MFT is benchmarked against the back-propagation learning algorithm (BP) on two different feature recognition problems: two-dimensional mirror symmetry and multidimensional statistical pattern classification. We find that while the two algorithms are very similar with respect to generalization properties, MFT normally requires a substantially smaller number of training epochs than BP. Since the MFT model is bidirectional, rather than feed-forward, its use can be extended naturally from purely functional mappings to a content addressable memory. A network with N visible and N hidden units can store up to approximately 4N patterns with good content-addressability. We stress an implementational advantage for MFT: it is natural for VLSI circuitry.</p>}}, author = {{Peterson, Carsten and Hartman, Eric}}, issn = {{0893-6080}}, keywords = {{Bidirectional; Content addressable memory; Generalization; Learning algorithm; Mean field theory; Neural network}}, language = {{eng}}, number = {{6}}, pages = {{475--494}}, publisher = {{Elsevier}}, series = {{Neural Networks}}, title = {{Explorations of the mean field theory learning algorithm}}, url = {{http://dx.doi.org/10.1016/0893-6080(89)90045-2}}, doi = {{10.1016/0893-6080(89)90045-2}}, volume = {{2}}, year = {{1989}}, }