Feedforward networks composed of units which compute a sigmoidal function of a weighted sum of their inputs have been much investigated. We tested the approximation and estimation capabilities of networks using functions more complex than sigmoids. Three classes of functions were tested: polynomials, rational functions, and flexible Fourier series. Unlike sigmoids, these classes can fit non-monotonic functions. They were compared on three problems: prediction of Boston housing prices, the sunspot count, and robot arm inverse dynamics. The complex units attained clearly superior performance on the robot arm problem, which is a highly non-monotonic, pure approximation problem. On the noisy and only mildly nonlinear Boston housing and sunspot problems, differences among the complex units were revealed; polynomials did poorly, whereas rationals and flexible Fourier series were comparable to sigmoids.
[1]
Kenneth D. Miller,et al.
Derivation of Linear Hebbian Equations from a Nonlinear Hebbian Model of Synaptic Plasticity
,
1990,
Neural Computation.
[2]
Robin Sibson,et al.
What is projection pursuit
,
1987
.
[3]
J. Friedman,et al.
Projection Pursuit Regression
,
1981
.
[4]
K. Miller,et al.
Ocular dominance column development: analysis and simulation.
,
1989,
Science.
[5]
G. Goodhill.
The Development of Topography and Ocular Dominance
,
1991
.
[6]
David J C MacKaytS,et al.
Analysis of Linsker's application of Hebbian rules to linear networks
,
1990
.
[7]
David J. C. MacKay,et al.
Analysis of Linsker's Simulations of Hebbian Rules
,
1990,
Neural Computation.
[8]
R. Linsker.
From basic network principles to neural architecture (series)
,
1986
.