A Generalist Neural Algorithmic Learner

Borja Ibarz, Vitaly Kurin, George Papamakarios, Kyriacos Nikiforou, Mehdi Bennani, Róbert Csordás, Andrew Joseph Dudzik, Matko Bošnjak, Alex Vitvitskyi, Yulia Rubanova, Andreea Deac, Beatrice Bevilacqua, Yaroslav Ganin, Charles Blundell, Petar Veličković
Proceedings of the First Learning on Graphs Conference, PMLR 198:2:1-2:23, 2022.

Abstract

The cornerstone of neural algorithmic reasoning is the ability to solve algorithmic tasks, especially in a way that generalises out of distribution. While recent years have seen a surge in methodological improvements in this area, they mostly focused on building specialist models. Specialist models are capable of learning to neurally execute either only one algorithm or a collection of algorithms with identical control-flow backbone. Here, instead, we focus on constructing a generalist neural algorithmic learner—a single graph neural network processor capable of learning to execute a wide range of algorithms, such as sorting, searching, dynamic programming, path-finding and geometry. We leverage the CLRS benchmark to empirically show that, much like recent successes in the domain of perception, generalist algorithmic learners can be built by "incorporating" knowledge. That is, it is possible to effectively learn algorithms in a multi-task manner, so long as we can learn to execute them well in a single-task regime. Motivated by this, we present a series of improvements to the input representation, training regime and processor architecture over CLRS, improving average single-task performance by over 20% from prior art. We then conduct a thorough ablation of multi-task learners leveraging these improvements. Our results demonstrate a generalist learner that effectively incorporates knowledge captured by specialist models.

Cite this Paper

BibTeX
@InProceedings{pmlr-v198-ibarz22a, title = {A Generalist Neural Algorithmic Learner}, author = {Ibarz, Borja and Kurin, Vitaly and Papamakarios, George and Nikiforou, Kyriacos and Bennani, Mehdi and Csord{\'a}s, R{\'o}bert and Dudzik, Andrew Joseph and Bo{\v s}njak, Matko and Vitvitskyi, Alex and Rubanova, Yulia and Deac, Andreea and Bevilacqua, Beatrice and Ganin, Yaroslav and Blundell, Charles and Veli{\v c}kovi{\' c}, Petar}, booktitle = {Proceedings of the First Learning on Graphs Conference}, pages = {2:1--2:23}, year = {2022}, editor = {Rieck, Bastian and Pascanu, Razvan}, volume = {198}, series = {Proceedings of Machine Learning Research}, month = {09--12 Dec}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v198/ibarz22a/ibarz22a.pdf}, url = {https://proceedings.mlr.press/v198/ibarz22a.html}, abstract = {The cornerstone of neural algorithmic reasoning is the ability to solve algorithmic tasks, especially in a way that generalises out of distribution. While recent years have seen a surge in methodological improvements in this area, they mostly focused on building specialist models. Specialist models are capable of learning to neurally execute either only one algorithm or a collection of algorithms with identical control-flow backbone. Here, instead, we focus on constructing a generalist neural algorithmic learner—a single graph neural network processor capable of learning to execute a wide range of algorithms, such as sorting, searching, dynamic programming, path-finding and geometry. We leverage the CLRS benchmark to empirically show that, much like recent successes in the domain of perception, generalist algorithmic learners can be built by "incorporating" knowledge. That is, it is possible to effectively learn algorithms in a multi-task manner, so long as we can learn to execute them well in a single-task regime. Motivated by this, we present a series of improvements to the input representation, training regime and processor architecture over CLRS, improving average single-task performance by over 20% from prior art. We then conduct a thorough ablation of multi-task learners leveraging these improvements. Our results demonstrate a generalist learner that effectively incorporates knowledge captured by specialist models.} }
Endnote
%0 Conference Paper %T A Generalist Neural Algorithmic Learner %A Borja Ibarz %A Vitaly Kurin %A George Papamakarios %A Kyriacos Nikiforou %A Mehdi Bennani %A Róbert Csordás %A Andrew Joseph Dudzik %A Matko Bošnjak %A Alex Vitvitskyi %A Yulia Rubanova %A Andreea Deac %A Beatrice Bevilacqua %A Yaroslav Ganin %A Charles Blundell %A Petar Veličković %B Proceedings of the First Learning on Graphs Conference %C Proceedings of Machine Learning Research %D 2022 %E Bastian Rieck %E Razvan Pascanu %F pmlr-v198-ibarz22a %I PMLR %P 2:1--2:23 %U https://proceedings.mlr.press/v198/ibarz22a.html %V 198 %X The cornerstone of neural algorithmic reasoning is the ability to solve algorithmic tasks, especially in a way that generalises out of distribution. While recent years have seen a surge in methodological improvements in this area, they mostly focused on building specialist models. Specialist models are capable of learning to neurally execute either only one algorithm or a collection of algorithms with identical control-flow backbone. Here, instead, we focus on constructing a generalist neural algorithmic learner—a single graph neural network processor capable of learning to execute a wide range of algorithms, such as sorting, searching, dynamic programming, path-finding and geometry. We leverage the CLRS benchmark to empirically show that, much like recent successes in the domain of perception, generalist algorithmic learners can be built by "incorporating" knowledge. That is, it is possible to effectively learn algorithms in a multi-task manner, so long as we can learn to execute them well in a single-task regime. Motivated by this, we present a series of improvements to the input representation, training regime and processor architecture over CLRS, improving average single-task performance by over 20% from prior art. We then conduct a thorough ablation of multi-task learners leveraging these improvements. Our results demonstrate a generalist learner that effectively incorporates knowledge captured by specialist models.
APA
Ibarz, B., Kurin, V., Papamakarios, G., Nikiforou, K., Bennani, M., Csordás, R., Dudzik, A.J., Bošnjak, M., Vitvitskyi, A., Rubanova, Y., Deac, A., Bevilacqua, B., Ganin, Y., Blundell, C. & Veličković, P.. (2022). A Generalist Neural Algorithmic Learner. Proceedings of the First Learning on Graphs Conference, in Proceedings of Machine Learning Research 198:2:1-2:23 Available from https://proceedings.mlr.press/v198/ibarz22a.html.

Related Material