Neural Networks Multiobjective Learning With Spherical Representation of Weights.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Multiobjective learning in artificial neural networks (ANNs) is typically a constrained optimization problem.
• Existing methods require significant computational resources to maintain constraints, limiting flexibility.
• The complexity of multiobjective optimization hinders the application of various nonlinear optimization techniques.

Purpose of the Study:

• To present a novel spherical representation for ANNs.
• To simplify the multiobjective learning problem by transforming it into an unconstrained optimization problem.
• To enable the use of diverse nonlinear optimization methods for ANN training.

Main Methods:

• Projecting ANN weights into a novel spherical space defined by radius and angles.
• Reformulating the constrained multiobjective learning problem as an unconstrained one.
• Applying standard nonlinear optimization techniques to the spherical representation.

Main Results:

• The proposed spherical representation simplifies the formulation and reduces computational effort for multiobjective learning.
• It yields more accurate estimates of the Pareto set compared to classical multiobjective approaches.
• The method effectively selects solutions from the Pareto set, outperforming state-of-the-art techniques on multiple datasets.

Conclusions:

• The spherical representation offers a computationally efficient and flexible alternative for multiobjective ANN learning.
• This novel approach enhances the accuracy of Pareto set estimation and solution selection.
• The method demonstrates superior performance over existing techniques in practical applications.