Title: False Discovery Rate Control and Simplification for Deep Neural Networks Using Knockoff-Based Approaches

Abstract: Deep neural networks have become a prevalent framework in machine learning, finding applications across numerous domains. However, these models typically encompass a vast number of parameters and input variables, many of which are irrelevant to the desired output or underlying goal. Such superfluous parameters and input variables not only elevate computational complexity but also impose additional computational burdens. To address this challenge, we leverage knockoff methods, which have demonstrated efficacy in managing false discovery rates within high-dimensional regression tasks. By integrating these methods with regularized neural networks, this study introduces three variable screening techniques designed to maintain control over false discovery rates: the one-layer filter, the multiple-layers filter, and the variable weight aggregation filter. Our results indicate that, compared to existing algorithms, the proposed methods deliver satisfactory performance.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC