Neural algorithmic reasoners are parallel processors. Teaching them sequential algorithms contradicts this nature, rendering a significant share of their computations redundant. Parallel algorithms however may exploit their full computational power, therefore requiring fewer layers to be executed. This drastically reduces training times, as we observe when comparing parallel implementations of searching, sorting and finding strongly connected components to their sequential counterparts on the CLRS framework. Additionally, parallel versions achieve (often strongly) superior predictive performance.
@article{arxiv.2307.04049,
title = {Parallel Algorithms Align with Neural Execution},
author = {Valerie Engelmayer and Dobrik Georgiev and Petar Veličković},
journal= {arXiv preprint arXiv:2307.04049},
year = {2024}
}
Comments
13 pages, 7 figures, Proceedings of the Second Learning on Graphs Conference (LoG 2023), PMLR 231