We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Persistent Message Passing

Abstract: Graph neural networks (GNNs) are a powerful inductive bias for modelling algorithmic reasoning procedures and data structures. Their prowess was mainly demonstrated on tasks featuring Markovian dynamics, where querying any associated data structure depends only on its latest state. For many tasks of interest, however, it may be highly beneficial to support efficient data structure queries dependent on previous states. This requires tracking the data structure's evolution through time, placing significant pressure on the GNN's latent representations. We introduce Persistent Message Passing (PMP), a mechanism which endows GNNs with capability of querying past state by explicitly persisting it: rather than overwriting node representations, it creates new nodes whenever required. PMP generalises out-of-distribution to more than 2x larger test inputs on dynamic temporal range queries, significantly outperforming GNNs which overwrite states.
Comments: 7 pages, 2 figures
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Social and Information Networks (cs.SI); Machine Learning (stat.ML)
Cite as: arXiv:2103.01043 [cs.LG]
  (or arXiv:2103.01043v1 [cs.LG] for this version)

Submission history

From: Petar Veličković [view email]
[v1] Mon, 1 Mar 2021 14:35:36 GMT (363kb,D)

Link back to: arXiv, form interface, contact.