Timezone: »

 
Poster
Understanding Attention and Generalization in Graph Neural Networks
Boris Knyazev · Graham Taylor · Mohamed Amer

Wed Dec 11 05:00 PM -- 07:00 PM (PST) @ East Exhibition Hall B + C #134
in Deep Learning -- Attention Models »

We aim to better understand attention over nodes in graph neural networks (GNNs) and identify factors influencing its effectiveness. We particularly focus on the ability of attention GNNs to generalize to larger, more complex or noisy graphs. Motivated by insights from the work on Graph Isomorphism Networks, we design simple graph reasoning tasks that allow us to study attention in a controlled environment. We find that under typical conditions the effect of attention is negligible or even harmful, but under certain conditions it provides an exceptional gain in performance of more than 60% in some of our classification tasks. Satisfying these conditions in practice is challenging and often requires optimal initialization or supervised training of attention. We propose an alternative recipe and train attention in a weakly-supervised fashion that approaches the performance of supervised models, and, compared to unsupervised models, improves results on several synthetic as well as real datasets. Source code and datasets are available at https://github.com/bknyaz/graphattentionpool.

Keywords: Attention Models · Deep Learning · Algorithms -> Relational Learning; Algorithms · Representation Learning; Deep Learning
[ Paper]  [ Poster]  [ Slides] 

Author Information

Boris Knyazev (University of Guelph / Mila)
Graham Taylor (University of Guelph)
Mohamed Amer (RobustAI)

More from the Same Authors