On the Complexity of Verifying Quantized GNNs with Readout

Artem Chernobrovkin, Marco Sälzer, François Schwarzentruber, Nicolas Troquard

Keywords: Graph neural networks, Complexity

Abstract:

In this paper, we introduce a logical language for reasoning about quantized graph neural networks (GNNs) with Global Readout. We then prove that verifying quantized GNNs with Global Readout is NEXPTIME-complete.
We also experimentally show the relevance of quantization in the context of ACR-GNNs.

Primary Area: Theory (e.g., control theory, learning theory, algorithmic game theory)

Submission Number: 20526