Reproduce “Deep Learning for Functional Data Analysis with Adaptive Basis Layers” on OpenResearch

About this paper

Junwen Yao, Jonas Mueller, Jane-Ling Wang

The paper introduces Adaptive Functional Neural Network (AdaFNN), a deep learning architecture that directly learns optimal basis functions for functional data in an end-to-end manner. AdaFNN consistently achieves superior predictive accuracy compared to traditional methods across various simulation and real-world functional data tasks, often using fewer basis functions.

Key insights

• Integrating basis function learning directly into a deep learning framework allows for the adaptive discovery of representations optimally tailored to the prediction task, moving beyond general variance capture.
• Learned basis functions can capture predictive signals that traditional variance-based methods like FPCA might miss, especially when those signals do not explain a large proportion of the input function's variation.
• Regularization for orthogonality and sparsity not only enhances model performance but also improves the interpretability of the learned functional representations by encouraging distinct, focused bases.

Summary by alphaXiv.