FreIE Low-Frequency Spectral Bias in Neural Networks for Time-Series Tasks
Published in IEEE International Conference on Data Mining (ICDM), 2025
We demonstrate that spectral bias—the tendency of neural networks to fit low-frequency signals first—is a universal trait in time-series prediction, not an architectural artifact. To mitigate this, we propose FreLE, an algorithm using frequency regularization to improve model generalization. Extensive experiments confirm FreLE’s effectiveness.