A new supervised learning algorithm for multilayered and interconnected neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Supervised learning in artificial neural networks typically relies on gradient-based optimization.
• Gradient methods can be computationally intensive and may face challenges with complex network architectures.

Purpose of the Study:

• To introduce a novel, gradient-free learning algorithm for multilayered and interconnected neural networks.
• To provide an alternative training method that avoids the computational complexities of traditional gradient descent.

Main Methods:

• The proposed algorithm combines an error backpropagation (EBP) method to generate fictitious teacher signals for hidden units.
• It then employs an exponentially weighted least squares (EWLS) method to determine the network's weight parameters.
• The approach is adapted for interconnected neural networks by mathematically arranging their descriptions.

Main Results:

• The EBP-EWLS algorithm successfully trains multilayered neural networks without gradient descent.
• The method is applicable to interconnected neural network structures.
• Simulation studies confirmed the efficacy and validity of the proposed learning technique.

Conclusions:

• The EBP-EWLS algorithm offers a viable gradient-free alternative for training neural networks.
• This method has the potential to simplify and accelerate the training of complex neural network models.
• Further research can explore its application in diverse machine learning tasks.