Fast training of multilayer perceptrons

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Training multilayer perceptrons (MLPs) using the error backpropagation algorithm is often slow and unreliable.
• Existing methods face challenges with convergence and can get stuck in local minima or experience paralysis.

Purpose of the Study:

• To introduce a novel, faster, and more reliable approach for training MLPs.
• To overcome the limitations of the traditional error backpropagation algorithm.

Main Methods:

• The proposed method combines iterative and direct solution techniques.
• It employs inverse transformation for linearizing nonlinear activation functions and direct matrix methods for output layer weight training.
• Gradient descent, the delta rule, and novel techniques are used for hidden layer weight training.

Main Results:

• The new approach demonstrates significantly faster training times compared to error backpropagation.
• Experimental results show comparable or superior recognition accuracy.
• The method effectively avoids local minima and paralysis, leading to more certain training outcomes.

Conclusions:

• The developed method offers a substantial improvement in MLP training efficiency and reliability.
• It provides a viable alternative to error backpropagation, achieving high accuracy with faster convergence.
• This approach enhances the practical applicability of MLPs in various problem domains.