Part of International Conference on Representation Learning 2025 (ICLR 2025) Conference
Isaac Reid, Kumar Dubey, Deepali Jain, William Whitney, Amr Ahmed, Joshua Ainslie, Alex Bewley, Mithun George Jacob, Aranyak Mehta, David Rendleman, Connor Schenck, Richard E Turner, René Wagner, Adrian Weller, Krzysztof Choromanski
When training transformers on graph-structured data, incorporating information about the underlying topology is crucial for good performance. Topological masking, a type of relative position encoding, achieves this by upweighting or downweighting attention depending on the relationship between the query and keys in the graph. In this paper, we propose to parameterise topological masks as a learnable function of a weighted adjacency matrix -- a novel, flexible approach which incorporates a strong structural inductive bias. By approximating this mask with graph random features (for which we prove the first known concentration bounds), we show how this can be made fully compatible with linear attention, preserving $\mathcal{O}(N)$ time and space complexity with respect to the number of input tokens. The fastest previous alternative was $\mathcal{O}(N \log N)$ and only suitable for specific graphs. Our efficient masking algorithms provide strong performance gains for image and point cloud data, including with $>30$k nodes.