Maximilien Burq

Currently

Senior Machine Learning Data Scientist & Technical Lead at Verily Life Sciences.

Technical Lead for a team of 4 building a suite of sensor-based digital measures and biomarkers for movement disorders. Our algorithms combine signal processing, statistics and deep learning to aggregate densely sampled time-series signals from wrist-worn wearable sensors into clinically relevant measures of motor impairment.

I always champion a problem-first approach: starting with the patient needs, we try off-the-shelf models and increase model complexity only where the performance benefit is clear. The model is a small part of the system; focus on simplicity, maintainability, clean data collection, rigorous validation, clean coding and testing.

Previously

  • [2014 - 2018] Ph.D. in Operations Research from MIT (Cambridge, MA, USA).

    • I had the privilege to work with Professors. Patrick Jaillet (MIT LIDS and ORC) and Itai Ashlagi (Stanford MS&E). My research focused on developing data driven matching algorithms in dynamic settings, with applications to kidney exchange programs.

  • [2017] Visiting researcher at Stanford University (CA, USA).
  • [2017] Research Scientist Intern at Lyft (CA, USA).
  • [2011 - 2014] MS in Applied Mathematics at École Polytechnique (France)

    • Focus on Probability theory, Optimisation, Operations Research and Machine learning. Ranked 1st (out of over 50000) on the nationwide entrance exam.

  • [2009 - 2011] Classes Préparatoires at Blaise Pascal (France)

    • Majors in Mathematics, Physics and Informatics.

Research Interests

  • Unsupervised and self-supervised representation learning for medical data. Data collection is costly in healthcare. Good labels even more so. Can we adapt recent advances in images and text representation learning to novel data modalities such as real-world evidence, sensor signals, -omics?
  • Interpretability, verification and validation. AI solutions need to be robust, interpretable and validated. How can we leverage attention mechanisms, traditional statistics, real world evidence / data to improve our confidence in the results, and shorten the time it takes to bring the solution to patients?