Transcending Tuning: Cross-domain Cross-platform Transfer Learning and Auto-tuning

• Applicant: Professor Dr.-Ing. Jeronimo Castrillon-Mazo
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 552689849

In the ever-evolving landscape of computing, the pursuit of performance across diverse domains and platforms has become a challenge. As applications and systems span multiple domains, from artificial intelligence to embedded systems, and platforms ranging from cloud computing to edge devices, the need for a unified approach to (auto-)tuning becomes evident. Tuning involves adjusting various parameters, settings, or configurations to improve the efficiency of a computation in terms of latency, resource utilization, or energy efficiency. In recent years, researchers have successfully leveraged machine learning techniques to improve the quality of auto-tuners. In this project, we intend to leverage recent advances in transfer learning to enable what we call transcending tuning, that is, an auto-tuning framework that can be adapted across applications domains and target computing systems. This is enabled by innovation on representations of code amenable to machine learning in compilers, adaptable and retargetable cost models, and a novel auto-tuning methodology based on transfer learning.

• DFG Programme: Research Grants