A New Lattice-Ladder Neural Network
1997 (English)Report (Other academic)
An idea to incorporate FIR or IIR filters in the common structure with Artificial Neural Network (ANN) is motivated by the need to process with ANN time-varying signals. Employment of static ANNs with input data windowing in time in this case is also possible, however experimental studies showed unsatisfactory results. In the common structures of filters and ANNs, IIR filters are more advantageous than FIR filters, as they require less coefficients and can handle long duration signals. However, non-guaranteed stability of IIR filters requires special attention. Both structures suffer from the long training time, because it is proportional to input data correlation which usually is big, e.g., in speech signals.We propose new ANN structure, which is based on the Lattice-Ladder realization of IIR filters incorporated as Multi-layer Perceptron (MLP) synapses. It is of small size because redundant ladder parts of filters are removed, has simple monitoring of filters stability, trains rapidly, and is easy expandable in filter orders. Last two properties are guaranteed because of orthogonality of signals inside of Lattice-Ladder filters. Comparison of proposed ANN structure with MLPs having full Lattice-Ladder or FIR filters in their synapses, and MLP with input data windowing in time in nonlinear system identification and speech signal prediction frameworks is given and confirms the new structure reasonability.
Place, publisher, year, edition, pages
Linköping: Linköping University Electronic Press, 1997. , 30 p.
LiTH-ISY-R, ISSN 1400-3902 ; 1981
Artificial neural networks, Reduced size lattice-ladder neural network
Cybernetik och informationsteori
IdentifiersURN: urn:nbn:se:liu:diva-55402ISRN: LITH-ISY-R-1981OAI: oai:DiVA.org:liu-55402DiVA: diva2:315980