Frequency-Domain Feature Learning
Overview
Frequency-domain feature learning studies how models represent patterns at different spatial or temporal frequencies, from periodic trends and rhythms to edges and fine textures. Common approaches use FFT, STFT, DCT, wavelets, or learned filter banks, combined with band-specific branches, spectral attention, adaptive filtering, or fusion with time- and spatial-domain features. These methods are applied to forecasting, image and audio tasks, EEG, graph learning, and inverse problems, while theoretical work examines why models learn some frequencies before others. A central tension is that low-frequency emphasis can capture global structure and improve stability but may lose fine detail; fixed spectral choices, phase and amplitude cues, and band balances do not transfer equally across tasks or domains.
Latest papers
195 papers12 papers in the last four weeks, up 9% on the four weeks before. 0.1% of all new papers.