Levenberg-Marquardt multi-classification using hinge loss function.

Area of Science:

• Deep Learning
• Optimization Algorithms
• Machine Learning

Background:

• Higher-order optimization functions like Levenberg-Marquardt (LM) offer improved generalization in deep learning.
• These methods face challenges with large datasets and complex problems like multi-view classification due to high processing time and training complexity.
• Finding global optima in multi-view classification is computationally expensive.

Purpose of the Study:

• To develop an efficient solution for Levenberg-Marquardt (LM)-enabled classification in multi-view scenarios.
• To implement hinge loss for the first time in conjunction with LM for multi-view classification.
• To address the limitations of processing time and training complexity in deep learning for large datasets.

Main Methods:

• Integration of hinge loss with the Levenberg-Marquardt (LM) optimization algorithm.
• Application to multi-view classification problems.
• Empirical validation across various multiclass classification challenges with diverse complexity and data sizes.

Main Results:

• The proposed method demonstrates faster convergence and improved performance compared to other loss functions.
• Significant reductions in training time were observed, particularly under time constraints.
• The approach achieved superior accuracy rates across all tested classification challenges.
• Outperformance was especially notable in scenarios with limited training time.

Conclusions:

• The combination of LM optimization and hinge loss offers an effective solution for efficient multi-view classification.
• This research highlights the critical impact of the interplay between optimization and loss functions on deep learning performance.
• The findings provide valuable insights for developing more efficient and accurate deep learning models for complex classification tasks.