I am currently an AI Research Resident at Intel Corporation. I am a part of the AI Product Group based in San Diego and working on the acceleration of Deep Neural Networks.
Before that I was an Inria postdoctoral researcher in the Arithmetic and Computing (AriC) team at LIP, ENS de Lyon.
I did my PhD in Computer Science at the LIP6 laboratory, University of Pierre and Marie Curie (now Sorbonne University). I was a member of the Performance and Quality of Numerical Algorithms (PEQUAN) team and took part in the ANR project “Metalibm“. My supervisors were Thibault Hilaire and Christoph Lauter. In September 2017 I defended my thesis entitled “Towards reliable implementation of digital filters” (slides and manuscript).
My current research concerns three main axes:
- Efficient rational approximations implemented on modern floating-point processors (collaboration with Jean-Michel Muller, CNRS)
- Guaranteed approximation of general numerical programs (collaboration with Eva Darulova, MPI)
- Design of optimal multiplierless digital filters on FPGA (collaboration with Martin Kumm, Kassel Unviersity and Silviu Filip, Inria)
Towards reliable and accurate software and hardware. Computer arithmetic approach.
My research aims to use and improve the methods of computer arithmetic for the design of reliable and accurate programs in the field of numerical computation. Programs rarely use exact arithmetic to perform their calculations, giving priority to the use of more efficient floating-point arithmetic. Embedded systems, in the search for better throughput and latencies, often fall back to fixed-point arithmetic. The use of these arithmetics results in computational errors that can have a serious impact on the result of the execution. In addition, choices about arithmetic (e.g. precision of computations and of data) inevitably influence the performance of the software and/or the target hardware cost. It is therefore important to analyze the computational errors and to determine software / hardware implementation parameters in an optimal and certified manner. Manual analysis of finite precision arithmetic is a tedious and non-trivial task, hence the need to incorporate this knowledge into code generators.
In my research I am passionate about the analysis of the numerical quality of algorithms and then the optimization of arithmetic parameters (precision, type of hardware operators, etc.) in order to guarantee given constraints a priori, especially error bounds. During my thesis I considered algorithms in the context of signal processing and control. Then, during my postdoctoral fellowship, I broadened the application area for general purpose numerical algorithms, where I focused on the implementation of mathematical functions, and optimal hardware design for digital filters on FPGA targets.