Knockoff methods cut DNN parameters by controlling false discovery rates

A team led by Huiqi Zhang has introduced a statistical framework to simplify deep neural networks by removing irrelevant parameters and input variables. Published on arXiv (2606.04404), the paper adapts knockoff methods—a proven technique for controlling false discovery rates in high-dimensional regression—to the context of DNNs. Specifically, they propose three variable screening algorithms: one layer filter, multiple layers filter, and variable weight aggregation filter. Each operates under strict false discovery rate control, ensuring that only truly relevant features are retained.

These methods leverage regularized neural networks to penalize unnecessary connections, then apply knockoff-based selection to identify which variables are actually contributing to the model's output. The result is a simplified architecture that reduces computational overhead without degrading predictive performance. While the paper does not provide benchmarks on common datasets, the authors report satisfactory results compared to existing pruning and feature selection techniques. For ML practitioners, this offers a principled way to automatically trim overparameterized models—a common pain point in deploying deep learning at scale.