Adaptive Manifold Concept with Regularized Autoencoders (AMRAE) for Effective Dimensionality Reduction

The Adaptive Manifold Concept with Regularized Autoencoders (AMRAE) algorithm is introduced as a novel dimensionality reduction technique that integrates manifold learning with autoencoders to capture the intrinsic geometry of high-dimensional data effectively. The study evaluates the impact of various adjustments and enhancements within the "Manifold Adjustment Box," including different regularization techniques, activation functions, and architectural choices, across diverse datasets. Key findings demonstrate that configurations such as Leaky ReLU Activation and Batch Norm Layer consistently improve accuracy, results highlight the flexibility and robustness of AMRAE. The results underscore the significance of AMRAE in addressing the limitations of traditional dimensionality reduction methods, with potential applications in various fields requiring high-dimensional data analysis. This paper provides a comprehensive summary of the objectives, methodology, results, and conclusions, showcasing AMRAE's efficacy in improving data representation and machine learning performance.