FourierQK: Filter Shape, Admissibility and the Leakage-Coverage Law

The study investigated frequency-collapse attention by replacing standard dot products with bandpass filters at learned frequencies. It found that success requires suppressing DC and Nyquist components, using a narrow bandwidth of about two bins at paragraph scale, and employing admissible filter shapes. These findings establish design rules for effective spectral attention in language modeling.

Open the full episode

More from AI Radio