Thanks for taking the time to check in!
What's this website: I try to give here a personal sense of who I am, both as an individual and as an engineer.
I am an AI research engineer at the Yale-LiGHT research lab (Laboratory for Intelligent Global Health & Humanitarian Response Technologies) which is shared between Yale and EPFL, Switzerland, where I am based.
Most of my professional experience is in machine learning research.
I graduated with an M.Sc. in Data Science, and before any technical knowledge, my studies taught me how to learn efficiently and imparted me a delight for learning.
I first think of myself as curious and interested. Accordingly, I love interdisciplinary projects － as you may notice in my experiences: medical diagnosis, digital education, financial fraud detection － for the opportunity to dive into a new discipline.
My peers have sometimes described me as thoughtful, dedicated and sweet.
Experience
AI Research Engineer － EPFL & Yale
2024
I am now working as a research engineer in a machine learning lab shared between EPFL and Yale.
I work on designing and implementing the no-code platform DISCO for federated and decentralized machine learning in collaboration with ICRC and WHO.
The platform allows training and using state-of-the-art models in low computational resource settings, via mobiles and personal computers. DISCO also enables
multiple entities to leverage confidential data to train models collaboratively without sharing their data between each other.
Concurrently, I am working in collaboration with the privacy and cryptography compagny Tune Insight on privacy-preserving training of Large Language Models.
The joint project aims at navigating the fine line between training the most helpful LLMs and preventing privacy leakage and sensitive data memorization during training.
EPFL － M.Sc. in Data Science
2020-2023
5.8/6
EPFL (Swiss Federal Institute of Technology, Lausanne) is a leading engineering school and is consistently ranked within the world top 10 universities in computer science.
University of California, Berkeley
－ Spring 2023 | Master's ThesisDesigned a new computer vision method to analyze error patterns in clinician misdiagnoses of skin cancer.
Oracle Labs, Zürich
－ Fall 2022 | InternshipImplemented a novel financial fraud detection method through graph neural networks. The method flags suspicious networks of transaction and provides interpretations for its decisions.
ML for Education Lab
－ 2020 - 2021 | Research ScholarFunded by a research grant, I worked alongside my studies at the ML4ED research lab, which leverages Machine Learning to improve digital learning. We utilized machine learning to stufy student's online learning methodology in flipped classroom settings to identify the most successful learning strategies.
EPFL － Bachelor in Computer Science
2017-2020
5.2/6
The bachelor covered the fundamentals of computer science: algorithms and data structures, multiple programming paradigms － object-oriented, imperative, functional － computer architecture, computer networks, etc.
KTH, Stockholm
－ 2019 - 2020 | Exchange yearProjects
Analyzing Idiosyncratic patterns of clinician errors in skin cancer diagnosis
In the context of my Master's thesis, I worked at the University of California, Berkeley on a new method to analyze error patterns in clinician misdiagnoses of melanoma skin cancer. To uncover idiosyncratic and systematic diagnostic errors, we relied on a clustering of computer vision image embeddings to analyze intra- and inter-clinician variations. The thesis received the maximal grade.
Financial fraud detection via graph neural networks on heterogeneous graphs
During a 6-month internship at Oracle Labs, we designed a novel graph neural network (GNN) method tailored to detect fraud cases in a financial transaction network. GNN allows a flexibility matching the diversity of forms frauds cases can take. As main innovations, the method detects whole transaction networks rather than individual nodes or transactions, and supports different types of vertices and edges, i.e., heterogeneous graphs.
Deep Learning for Automated Detection of Tuberculosis from Lung Ultrasound
A 6-month research project, in which I leveraged deep learning to diagnose TB from lung ultrasounds (LUS). Ultrasound probes's practicality makes it a promising alternative. TB is nowadays only present in regions where treatment can't be afforded. However, LUS interpretation is notoriously complex and deep learning may prove pivotal in TB diagnosis. Our model proved significantly better than clinician-handcrafted features.
Machine Learning to improve digital learning and student retention
Funded by an EPFL grant during 12 months, I worked alongside my studies in a lab leveraging Machine Learning to improve education. The first project aimed at improving early prediction of student dropout in online courses. In a second time, I worked on a deep learning framework to handle irregular time series, ubiquitous in digital education and currently requiring manual engineering. Both projects resulted in publications.
Trip planner accounting for delays and successful connection probabilities
The aim of the project is to leverage the Swiss public transport data to implement a journey planner in the area of Zurich. Notably, our trip planner assesses the success rate of each connection, letting the user choose the confidence level of his or her journey. The data, stored as a Hive table, was processed using PySpark and the algorithm was implemented in Python.
Chess insights through visualizations of one million chess games
By collecting and parsing more than a million chess games from Lichess, we set our goal to draw insights and convey them intuitively through interactive data visualizations using D3.js. Starting from a visual history to the game basics, we show the importance of the chessboard center and analyze the effectiveness of the different chess openings in the dataset. We also let users watch any game in the dataset to follow their own interests.
Song2vec: mining music listening history for meaningful song embeddings
Inspired by the word2vec method, we utilized the music listening history of 1000 users, composed of 20 million listens, to extract information found in songs' listening chronology. Such method is essential in music recommendation. Interestingly Karma Police - Radiohead + Metallica = Where Is My Mind? In other words, if Karma Police had been produced by Metallica rather than Radiohead, our closest result was Where Is My Mind? by The Pixies.
Movie recommender systems for large scale data with Apache Spark
Throughout this course project, we implemented multiple recommender systems of increasing complexity in Apache Spark and Scala. The project is organized as successive milestones to practice wielding Scala and Spark to handle large scale data.
Machine Learning competition of road segmentation from satellite images
We designed and optimized a model to discriminate roads from satellite images. We implemented a U-Net model with TensorFlow Keras, producing a 2D-grid of pixel labels. By leveraging heavy data augmentation during training and majority voting on image transformations during inference, our model reached an F1 of 90% and an accuracy of 94.5%. We used an Adam optimizer and a focal Tversky loss to address the class imbalance.
Interests
Hiking
Picture from the top of Mont Charvin, France.
I am usually up in the mountains whenever time and weather permitting. I grew up in the Alps, and mountains are typically the first thing I would miss in a new location.
Cycling
A farewell gift from a dear friend, showing Bikie my dear bike.
Being able to commute by bike is one of the simple things I appreciate and treasure.
Linguistics
My name, Julien, spelled with the International Phonetic Alphabet.
Learning Vietnamese in a self-taught manner created a deep interest in linguistics and led me to audit a linguistics class during my stay at UC Berkeley.
I am fascinated by the diversity found in languages, especially in phonetics and writing systems.
Arts and craft
These flowers were painted as a token of gratitude for some friends.
I dabble in creative crafts, mostly drawing and watercolor. I've recently tried linocut printmaking and loved it. Furthermore, I enjoy creative programming tasks, such as data visualization or designing this website.
Reading
A sunny morning at the farmer's market, snacking on strawberries while reading.
Whether in transit or waiting in line, I always have a book on me. A good book with a drink in a cute location is one of the ways I savor life. I read mostly fiction, essays and whatever my friends recommend. Additionally, I'm very happy to discuss the pros and cons of an e-reader.)
Learning
I deeply enjoy diving into subjects I find compelling that are usually outside my field of study, linguistics being one of them. I also read about psychology － I took multiple classes of social psychology during my studies － architecture and city planning, or art history. I am curious and enjoy learning more without any pretention. I rely mostly on books and Wikipedia. As such, the literature review is of my favorite parts of a project.
Traveling
As common as it sounds, I like exploring new places, witnessing different architecture styles, new landscapes and urbanism, trying out local cuisines and learning about places' history. I strive to minimize my airplane carbon footprint and I travel mostly by train. I recently traveled all over the US for two months via night trains.
Cooking
Pretty produce at Jean-Talon Market in Montréal.
While I don't consider myself a very talented cook, I always enjoy the time I spend cooking. Especially improvising a yummy dish with whatever is in the fridge. I appreciate the creativity, the educated intuitions and satisfying results. I also really like eating.
Publications
Proficiency is Associated with Higher Idiosyncratic Biases in Medical Image Perception.
† Equal Contributions.
Working paper.
Ripple: Concept-Based Interpretation for Raw Time Series Models in Education.
Proceedings of the AAAI Conference on Artificial Intelligence, 37(13), 15903-15911.
Can Feature Predictive Power Generalize? Benchmarking Early Predictors of Student Success across Flipped and Online Courses.
Proceedings of the 14th International Conference on Educational Data Mining.