OxCSML seminar this week: Antonio Orvieto on Friday 2-3pm
Hai Dang Dau
From: Yee Whye Teh <y.w...@stats.ox.ac.uk>
Date: Mon, Nov 27, 2023 at 11:24 PM
Subject: [ML] OxCSML seminar: Antonio Orvieto on Friday 2-3pm
To: m...@maillist.ox.ac.uk <m...@maillist.ox.ac.uk>, oxc...@googlegroups.com <oxc...@googlegroups.com>
This Friday we will have Antonio
Orvieto visiting and giving the OxCSML seminar at 2-3pm. He would be happy to meet with individuals here in Oxford on Friday. Please let me know if you are interested in meeting him and let me know when you are available to meet him.
OxCSML Seminar
Dec 1, Friday 2-3pm
Small lecture theatre, Department of Statistics, and on zoom
Speaker: Antonio Orvieto, ELLIS Institute Tübingen
Title: Long-range reasoning on graphs without attention
Abstract: Graph neural networks based on iterative one-hop message-passing have been shown to struggle in harnessing information from distant nodes effectively. Conversely, graph transformers allow each node to attend to all other nodes directly, but suffer from high computational complexity and have to rely on ad-hoc positional encodings to bake in the graph inductive bias. In this talk, we present a new architecture to reconcile these challenges. Our approach stems from the recent breakthroughs in long-range modeling provided by deep state-space models on sequential data (S4, LRU, etc..). For a given target node, our model aggregates nodes at different distances and uses a parallelizable linear recurrent unit (LRU) over the chain of distances to provide a natural encoding of its neighborhood structure. With no need for positional encoding, we empirically show that the performance of our model is competitive compared with that of state-of-the-art graph transformers on various benchmarks, at a drastically reduced computational complexity. In addition, we show that our model is theoretically more expressive than one-hop message-passing neural networks.
Zoom: https://zoom.us/j/91238172693?pwd=dzhIQUx1MG9XZHk1R3QvbUZrYXRGUT09