Seven papers authored by Computer Science researchers from Warwick have been accepted for publication at the 37th Conference on Neural Information Processing Systems, the leading international venue for machine learning research, which will be held on 10-16 December 2023 in New Orleans, Louisiana, USA:

• EV-Eye: Rethinking High-frequency Eye Tracking through the Lenses of Event Cameras, by Guangrong Zhao, Yurun Yang, Jingwei Liu, Ning Chen, Yiran Shen, Hongkai Wen, and Guohao Lan
• Fully Dynamic k-Clustering in Õ(k) Update Time, by Sayan Bhattacharya, Martin Costa, Silvio Lattanzi, and Nikos Parotsidis
• Initialization Matters: Privacy-Utility Analysis of Overparameterized Neural Networks, by Jiayuan Ye, Zhenyu Zhu, Fanghui Liu, Reza Shokri, and Volkan Cevher
• Learning a Neuron by a Shallow ReLU Network: Dynamics and Implicit Bias for Correlated Inputs, by Dmitry Chistikov, Matthias Englert, and Ranko Lazic
• On the Convergence of Shallow Transformers, by Yongtao Wu, Fanghui Liu, Grigorios Chrysos, and Volkan Cevher
• Towards Data-Agnostic Pruning At Initialization: What Makes a Good Sparse Mask? by Hoang Pham, The Anh Ta, Shiwei Liu, Lichuan Xiang, Dung Le, Hongkai Wen, and Long Tran-Thanh
• Towards Unbounded Machine Unlearning, by Meghdad Kurmanji, Peter Triantafillou, and Eleni Triantafillou

We are delighted to announce that Peter Kiss, a PhD student in the Theory and Foundations Research Division, has received the best student paper award at European Symposium on Algorithms (ESA) 2023, for his joint work with Joakim Bilkstad for the paper: "Incremental (1-eps)-approximate dynamic matching in O(poly(1/eps)) update time". The paper considers the problem of maintaining a large matching in a graph that is undergoing a sequence of edge insertions. They present an algorithm for this fundamental problem in dynamic graph algorithms, which has near-optimal approximation ratio and an update time that does not grow at all with the size of the input and is also polynomial in 1/\eps (the error parameter). In addition, their approach is simpler than previous algorithms on the same problem that achieved weaker guarantees.

Mustafa Yasir, a former Warwick Department of Computer Science student who graduated in Summer 2023, wrote up and presented an academic paper on the work carried out as part of his third year project. The paper was accepted to the 3rd Annual Workshop on Graph Learning Benchmarks at KDD 2023, and was presented in California by Mustafa.

Mustafa's third year project idea, supervised by Dr Long Tran-Thanh and titled 'Extending the Graph Generation Models of GraphWorld', started whilst he was interning at Google last summer. Mustafa contacted some researchers at the company working in the Graph ML space, to ask for any relevant project ideas. He bumped into a team who had just published GraphWorld: a tool to change the way Graph Neural Networks are benchmarked, by creating synthetic graph datasets through graph generation models – as opposed to using real-world datasets that are limited in their generalisability and present a major issue facing the field of Graph Learning.

However, since GraphWorld only used a single graph generation model in this process, Mustafa integrated two additional models with the system, ran large-scale GNN benchmarking experiments with these models and published his code to Google’s official GraphWorld repository. The project provides a significant advancement to researchers across the field looking to benchmark models and guide the development of new architectures.

Dr Long Tran-Thanh commented:

What Mustafa and the GraphWorld team has been working on is very important for the machine learning and AI research communities. In particular, there has been a vocal criticism against the whole field that most models are trained on the same public datasets (e.g., ImageNet, MNIST, etc), therefore are not diverse enough. One way to mitigate this issue is to generate realistically looking synthetic data. This need is especially of importance in within the graph learning community. GraphWorld’s aim is to address this exact problem by creating a powerful and convenient tool that can generate a diverse set of graphs, ranging from large social network-style graphs to molecule-inspired ones. Joining this project with the Google researchers is a huge opportunity for Warwick students to participate in a very impactful project.