Morning Edition · Thursday, July 23, 2026Published at 2:03 AM EDT · New York

Researchers Propose Subquadratic Operators That Keep a Global View of Multi-Dimensional Data

Input-dependent long convolutions aim to give images and video attention-like receptive fields without attention's quadratic cost or the rasterization recurrent models require.

A paper on native multi-dimensional subquadratic operators targets a persistent tradeoff in efficient architectures. Subquadratic alternatives to attention tend to compromise on multi-dimensional data: standard convolutions lack a global re…