perceiver-latent-array-decouples-compute-from-input-length

Summary

Instead of letting every element in the input interact with every other element, which gets explosively expensive as inputs grow, the system compresses attention to a small fixed set of learned queries that "read" the input of any size and then iteratively refine the result. The practical upshot is that processing cost stays roughly constant no matter how long the input is, so the system can handle very long sequences like high-resolution images or extended audio without a proportional jump in compute.