SUMMARY

In the mid 90s, I attempted to fundamentally rethink the design and optimization of computer systems to considerably speed up modeling of object recognition in cortex using my semiconductor neural networks. During this long term research endeavor, I proposed and developed multiple theoretical and practical techniques to build faster and more power efficient devices based on my interdisciplinary background in physics, electronics and artificial intelligence. I managed to effectively combine collective participation of multiple users with empirical and statistical program and architecture analysis and optimization, machine learning and run-time adaptation. Eventually, this technology has become a mainstream trend to build efficient computer systems resulting in the world’s first machine learning enabled self-tuning compiler (MILEPOST GCC) and public collective tuning center (cTuning). It also fundamentally revisits and systematizes research methodology, enables reproducibility of the research results, and is being actively used and extended by academia and industry (IBM, Synopsys, Intel, Google and others). This technology have been included in the HiPEAC 2012-2020 research roadmap, referenced in patents and conferences/journals such as PLDI, MICRO, CGO, CASES, HiPEAC, IJPP, TACO, and in world-wide press-releases from IBM. In 2010-2011, I have been invited to help establish new Intel Exascale Lab in France as the head of application characterization and optimization group. I also regularly give invited talks about my research at companies and universities including Intel