Over the recent year, Neural Architecture Search (NAS) has attracted a lot of attention. While being able to automate the discovery of better performing neural architectures than hand-crafted ones, it comes at a great price, requiring thousands of GPU hours to perform the search. The Zero Cost NAS competition challenges the participants to design efficient proxies for NAS, using negligible computational resources to evaluate neural architectures.
In collaboration with the AutoCAML team at Samsung AI Cambridge (led by Dr. Hongkai Wen), our research students, Lichuan Xiang and Youyang Sha, proposed new zero-cost NAS metrics that exploit the compressibility of neural networks. Our metrics are extremely efficient to run (reducing search cost from weeks/days to minutes), and achieves impressive results across multiple search spaces and datasets. In the competition, our teams won both the 1st and 2nd places (using different scoring functions), and the performance gap with the 3rd winning team is almost 2x. Checkout our poster here.
Oral Evidence to the House of Lords on Telephone Frauds and Countermeasures
On 23rd June 2022, Professor Feng Hao of the Systems and Security research theme was invited as one of the two expert witnesses to give oral evidence to the Fraud Act 2006 and Digital Fraud Committee appointed by the House of Lords at Parliament on trends of telephone frauds and the landscape of counter-fraud technologies. This is related to an ongoing EPSRC project, led by Professor Feng Hao (PI) from the Department of Computer Science and Dr Adrian von Mühlenen (co-I) from the Department of Psychology, the University of Warwick. In this project, the research team have been investigating a cost-effective solution to combat caller ID spoofing, a technique commonly used by fraudsters and scammers to pretend to call from trusted sources (e.g., banks, HMRC) as part of social engineering attacks. A transcript of the oral evidence session is published on the Parliament website.
HPC Research Accepted for Publication at the ICS 2022 Conference
Two papers by researchers at the Department of Computer Science have been accepted to the 36th ACM International Conference on Supercomputing ICS 2022 to be held on the 28-30th of June this year. ICS is one of the most prominent and revered conferences in High Performance Computing, highly regarded by the HPC community for publishing leading-edge research in this area. The two papers accepted are:
- High Throughput Multidimensional Tridiagonal System Solvers on FPGAs (Preprint) by Kamalavasan Kamalakkannan and Gihan Mudalige at Warwick, together with Istvan Reguly (PPCU) and Suhaib Fahmy (KAUST).
- Clairvoyant: A Log-Based Transformer-Decoder for Failure Prediction in Large-Scale Systems by Khalid Alharthi and Arshad Jhumka at Warwick, together with Sheng Di, Franck Cappello at Argonne National Laboratory. Preprint. The ACM ICS2022 full program can be found here.
The Research Excellence Framework (REF) is the UK’s system for assessing the quality of research in the country's higher education institutions.
The results of the 2021 REF rank Warwick Computer Science 4th out of 90 UK computing departments. This cements our position as one of the top Computer Science departments in the UK, a position we have held for some time under different assessment methodologies.
Winner of the Faculty of Science, Engineering and Medicine Post-Doctoral Research Prize 2022
Gunduz Vehbi Demirci has been awarded with the Faculty of Science, Engineering and Medicine Post-Doctoral Research Prize 2022 for his paper jointly with Prof. Hakan Ferhatosmanoglu, "Partitioning sparse deep neural networks for scalable training and inference", published in the Proceedings of the ACM International Conference on Supercomputing (ICS '21) (DOI: https://doi.org/10.1145/3447818.3460372).
Training large-scale deep learning models is notoriously difficult. Gunduz develops a highly parallel solution to scale training of sparse deep learning models, which is combined with a novel combinatorial optimisation built on a hypergraph partitioning model, reducing parallelisation overheads and achieving computational balance among processors. An end-to-end software solution is released, enabling competing with big tech companies that have access to large infrastructures and datasets.
The work is summarised in a paper accepted by the 2021 ACM International Conference on Supercomputing, which is a premier conference in high-end systems. The research output will have a great potential to bring significant practical impact in long term as developing such comprehensive solutions takes time and is typically achieved only within large groups.
Four papers accepted to STOC 2022
We are pleased to report that members of the department's Theory and Foundations research theme have had four papers accepted to the 54th ACM Symposium on Theory of Computing (STOC 2022), the ACM flagship conference in theoretical computer science. The papers are:
- "Deterministic Massively Parallel Connectivity" by Sam Coy and Artur Czumaj.
- "Improved Approximation Guarantees for Shortest Superstrings using Cycle Classification by Overlap to Length Ratios", by Matthias Englert, Nicolaos Matsakis, and Pavel Veselý.
- "Hypercontractivity on High Dimensional Expanders" by Tom Gur, Noam Lifshitz, and Siqi Liu.
- "Worst-Case to Average-Case Reductions via Additive Combinatorics" by Vahid R. Asadi, Alexander Golovnev, Tom Gur, and Igor Shinkar.