Multilayer Potts perceptrons with Levenberg-Marquardt learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Traditional perceptrons use sigmoid-like transfer functions.
• Multilayer perceptrons are widely used for function approximation.
• Limitations exist in the flexibility of traditional perceptron transfer functions for complex nonlinear mappings.

Purpose of the Study:

• Introduce multilayer Potts perceptrons (MLPotts) for data-driven function approximation.
• Extend the function space coverage of traditional multilayer perceptrons.
• Enhance the flexibility of nonlinear mappings in neural networks.

Main Methods:

• Developed a novel Potts perceptron architecture.
• Utilized a K-state transfer function generalized from sigmoid functions.
• Employed the Levenberg-Marquardt (LM) algorithm for MLPotts learning.
• Organized MLPotts networks for high-dimensional input translation to postnonlinear projections.

Main Results:

• MLPotts networks theoretically cover the function space of traditional multilayer perceptrons.
• Weighted Potts perceptrons offer more flexible postnonlinear functions.
• MLPotts learning using the LM method significantly improved function approximation accuracy.
• Demonstrated superior performance over traditional supervised learning methods for multilayer perceptrons.

Conclusions:

• MLPotts represent a powerful new model for data-driven function approximation.
• The enhanced flexibility of Potts perceptrons leads to improved learning performance.
• The combination of MLPotts and the LM method offers a significant advancement in supervised learning.