Monthly
288 pp. per issue
6 x 9, illustrated
ISSN
0899-7667
E-ISSN
1530-888X
2014 Impact factor:
2.21

Neural Computation

January 1993, Vol. 5, No. 1, Pages 115-131.
(doi: 10.1162/neco.1993.5.1.115)
© 1993 Massachusetts Institute of Technology
Training Periodic Sequences Using Fourier Series Error Criterion
Article PDF (860.48 KB)
Abstract

Training a network to learn a set of periodic input/output sequences effectively makes the network learn a mapping between amplitudes and phases in Fourier space. The spectral backpropagation (SBP) training algorithm is a different way of doing this task. It measures the Fourier series components of the output error sequences and minimizes the total spectral energy as an adaptation criterion. This approach can train not only the weights but also time delays associated with the interconnects. Furthermore, the cells can have finite bandwidth via a first-order low-pass filter. Having adaptable time delays gives the SBP algorithm a powerful way to control the phase characteristics of the network.