profile picture

Maximiliana Behnke

Deep Learning R&D Engineer

Languages

  • Polish (Native)
  • English (Professional)
  • Japanese (Basic)

Interests

  • Collecting physical media
  • Video games
  • Comics & Animation

Career Profile

Hi! My name is Maxi, a Deep Learning R&D Engineer at Intel. I graduated from a PhD programme at Centre for Doctoral Training programme in Pervasive Parallelism. My research area is exploring and optimising large-scale deep learning methods applied to machine translation and language models. I am interested in natural language processing, inference optimisation and data mining. I am seeking a research-oriented position to leverage expertise in advancing AI technologies and driving innovative solutions.

Education

Doctor of Philosophy

2018 - 2022
The University of Edinburgh, United Kingdom

Programme Centre for Doctoral Training in Pervasive Parallelism

Supervisors Kenneth Heafield, Roman Grundkiewicz

Thesis Structural Pruning for Speed of Neural Machine Translation

Passed viva with no corrections.

Master of Science by Research

2017 - 2018
The University of Edinburgh, United Kingdom

Programme Centre for Doctoral Training in Pervasive Parallelism

Supervisors Kenneth Heafield, Rico Sennrich

Thesis Improving Neural Machine Translation Training via Weighted Objectives

Graduated with distinction.

Bachelor of Engineering

2013 - 2017
Adam Mickiewicz University, Poland

Specialization Information and Internet Technologies

Thesis Computer vision and machine learning for automatic extraction of death notices

  • Average grade: 4.63/5
  • Thesis grade: 5/5 Viva grade: 5/5
  • Graduation ranking: 4/103 (top 5%)

Scholarships Rector’s Scholarship for Best Students 2014-2017

Experiences

Deep Learning R&D Engineer

2022 - Present
Intel, Edinburgh, United Kingdom

Heterogenous Systems Technologies team in Software and Advanced Technology group.

  • Leading cross-corporative efforts optimizing AI models for Intel’s CPUs.
  • Benchmarking and investigating inefficiencies in AI workloads deployed with oneDNN.
  • Prototyping an inference-focused AI accelerator.

Consultant

2020, 2021
Intel, Edinburgh, United Kingdom

Working on efficient machine translation shared tasks and pruning solutions.

Tutor & Demonstrator

2018 - 2020
The University of Edinburgh, Edinburgh, United Kingdom

Part time teaching vacancies.

  • Introductory Applied Machine Learning
  • Machine Learning Practical
  • Machine Learning Base — providing inplace support for students doing various ML courses.

Intern

September 2018 - December 2018
Ebay, Aachen, Germany

Assigned to Machine Translation team.

Research Associate

March 2017 - August 2017
The University of Edinburgh, Edinburgh, United Kingdom

Project TraMOOC – Translation for Massive Open Online Courses

Collaborating Universities HU Berlin, DCU, IURC, Radboud University, Tilburg University

The collaborative project aiming at providing reliable machine translation for online courses.

Intern

July 2015 - December 2016
Samsung R&D Institute, Poznań, Poland

Branch: Artificial Intelligence, Group Natural Language Processing, Team Dialog Translation

Assigned to Machine Translation sub-team.

Publications

  • Pruning Neural Machine Translation for Speed Using Group Lasso
    • Behnke, M., Heafield, K.
    The 6th Conference on Machine Translation (WMT21)
  • Efficient Machine Translation with Model Pruning and Quantization
    • Behnke, M., Bogoychev, N., Aji, A. F., Heafield, K., Nail, G., Zhu, Q., Tchistiakova, ..., Grundkiewicz, R.
    The 6th Conference on Machine Translation (WMT21)
  • Losing Heads in the Lottery: Pruning Transformer Attention in Neural Machine Translation
    • Behnke, M., Heafield, K.
    2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)
  • Edinburgh's Submissions to the 2020 Machine Translation Efficiency Task
    • Bogoychev, N., Grundkiewicz, R., Aji, A. F., Behnke, M., Heafield, K., Kashyap, S., ... & Chudyk, M.
    The 4th Workshop on Neural Generation and Translation (WNGT 2020)
  • Improving Machine Translation of Educational Content via Crowdsourcing
    • Behnke, M., Miceli-Barone, A. V., Sennrich, R., Sosoni, V., Naskos, T., Takoulidou, E., ... & Georgakopoulou, P.
    11th edition of the Language Resources and Evaluation Conference (LREC 2018)

    Projects

    Efficiency Shared Task at Conference on Machine Translation 2021 - Led the University of Edinburgh's team and worked on preparing the data, training and fine-tuning transformer models for submissions.
    Marian - Contributing to the development of Marian - a fast C++ Neural Machine Translation toolkit.
    Parallel Corpora Crawler - The parallelizable system that crawls and extracts parallel sentences in two specified languages given a list of domains.
    Gazette Extractor - The system that extracts death notices and obituaries from 20th century Polish newspapers by applying solutions from computer vision and machine learning. Done before neural networks were the main research direction to go.
    Playstation Manual Similarity - Applied information retrieval techniques to corpora made of Playstation video-game manuals to compare them with each other.
    MangaEditor - A computer-vision program that automatically cleans speech bubbles in comics, tested and applied to Japanese and English languages.

    Skills & Proficiency

    Python

    C++

    C

    Linux/Bash

    Git

    Pytorch

    Tensorflow

    Marian