Eyyüb Sari

Senior Machine Learning Engineer at Meta

Feb 2022 – Present

Machine Learning Research Engineer at Huawei Technologies

Oct 2018 – Feb 2022

• Conducted research and development for deep learning model compression based on requirements from Huawei's product teams. - Quantization (e.g., binary, ternary, 8-bit) - Pruning (e.g., block, channel, node, grid) - Knowledge distillation - Accelerated training • Trained Computer Vision (e.g., ResNet), Natural Language Processing (e.g., LSTM, Transformer), and Speech models using: - Distributed training - Gradient accumulation - Hyper-parameter tuning - Mixed-precision training - Data and model parallelism • Built internal model compression framework in PyTorch used in a dozen of research projects and transferred to product departments (Cloud, Consumer) • Delivered a compressed CNN model for Face Recognition using 8-bit quantization and pruning to the product. • Proactively developed an on-device 16-bit integer accelerated training application for digit recognition on Android using Kotlin, C++, and ARM NEON intrinsics. • Contributed to compression of ASR Voice Assistant for cars and mobile phones using quantization, pruning, and knowledge distillation. (TensorFlow, TF Lite, Horovod) • Researched the role of Batch Normalization in low-bit quantized networks. • Designed Differentiable Mask Pruning, a novel all-scenario task-agnostic method for pruning DNNs while training, removing the usually large finetuning time required. The work was patented, published in CRV 2020, and delivered to product departments (Cloud, Consumer). (Presentation: https://bit.ly/3juoBxx) • Performed research on integer-only recurrent neural networks with support for modern operations (e.g., normalization, attention). Implemented an integer-only LSTM achieving up to 2x inference speedup on a Huawei P20 Pro, using a custom PyTorch Mobile backend with C++ and ARM NEON intrinsics kernel. The work was patented and transferred to the consumer product line. Awards: - Director's Individual Award 2019 - CBG Software Director's Team award (2019, 2020)

Research Engineer Intern (Deep Learning) at Deeplite

Apr 2018 – Sep 2018

- Built an end-to-end platform for compressing neural networks, using PyTorch, aimed at bringing AI on device. - Explored and implemented works from the literature of optimizing DNNs for size, accuracy and computational cost. - Explored the optimization design space using Reinforcement Learning to satisfy given constraints eg. FLOPs. - Compression techniques explored: Pruning, Quantization, Dynamic Fixed-point, Multiplier-free, Knowledge Distillation, and Binary Neural Networks.

Backend developer - freelance at MedPics

Dec 2016 – Feb 2017

• Building an admin panel for MedPics's channels (Node.js, TypeScript, Mongo) • Code refactoring to take advantage of TypeScript's type system

Software Engineer Intern at Numalis

Dec 2015 – Mar 2016

• Developing an expression-based scripting language that compiles itself to a proprietary programming language (OCaml) • Making an OCaml binding for a C++11 proprietary library (OCaml, C++11) • Refactoring OCaml-based codebase (modern build system, lastest OCaml features)

Software Engineer Intern at Mobixio

Aug 2014 – Dec 2014

I prototyped an interactive map and a platform for transports(bus, tramway...). The user can see in real time, the transports moving on the map with useful informations, and receive notifications. The operator can manage notifications and the shipping. We designed a generic specification for transport using Open Data. • Speeding up the data processing codebase by 40% within the first week • Building a viable prototype of the platform (Javascript, Python, Flask) • Designing a data format specification for public transports based on GeoJSON

Master's degree, Computer Science at EPITECH - European Institute of Technology

2016 – 2018

Computer Science at Dublin City University

2016 – 2017

Bachelor's degree, Computer Science at EPITECH - European Institute of Technology

2013 – 2016

Bac S - spé ISN at Lycée Joffre - Montpellier

2010 – 2013