About Me
I am a sociable, meticulous student with an excellent ability to multi-task and demonstrate great resistance to stress.
I have skills in the fields of finance, public relations, and computer science.
I am 25 years old and currently in my third year of Ph.D. in computer sciences at Rennes 1 University and Inria laboratory.
My thesis is about the Automatic Construction of Explanations for AI Models.
My supervisors are Christine Largouët and Luis Galárraga from the
LACODAM team.
Due to recent advances in AI and in particular in deep learning, models are becoming more and more difficult to trust since their success is often due to their high complexity.
Relying on the answers of a black box can be an issue for technical, ethical, and legal reasons.
The purpose of my Ph.D. is to extend actual interpretable methods for machine learning methods to explain the inner mechanisms of a complex machine model.
During the first part of my Ph.D.,
I studied the generation of post-hoc local explanation and proposed two frameworks to select a priori the best surrogate model for tabular and text data.
Next, I am searching to collaborate with international researchers from another laboratory or company to study more in detail the impact of the surrogate model
on human understandability.
My projects are about integrating the user in the loop for interpretability (or what I want to call Activate Interpretability Learning)
or conducting user studies to evaluate the impact of the chosen explanation model.
If you are interested in collaboration, don't hesitate to send an email!
Work experience
2022 -
PhD student visiting Niels van Berkel at Aalborg University in Danemark for research collaboration
After having developped methods such as
APE to evaluate a priori whether a linear explanation is suitable to approximate locally a black box model,
I am currently visiting the associate Professor Niels van Berkel to conduct user studies in order to assess the users understanding depending on the explanation module.
Our first step is to compare the understandability of linear, rule-based and counterfactual explanations. In a second round, we will evaluate the impact of the user experience in his understanding.
2020 -
PhD student supervised by Christine largouët and Luis Galárraga in the domain of interpretability
The recently approved GDPR contemplates the right of individuals to contest decisions made on their personal data.
That encompasses decisions made by algorithms. To select the most suitable explanation for a use case taking into account criteria such as fidelity,
complexity, scope, semantics, and the target user. Such a task is far from trivial because it requires a knowledge of the guarantees of the different
explanation methods, as well as an analysis of the context in which the explanation will be delivered.
This task is time-consuming and unfeasible for non-technical users. We argue, however, that it can be automated.
2020
6 months internship in Lacodam team supervised by Luis Galárraga and Christine largouët
I complete my research master's degree with a second internship in a research laboratory.
This internship was part of the FABLE project (that leads to my PhD thesis.)
The purposes was to think and solve the question: "when are anchors based explanation not a good idea?"
I firstly studyied
Anchors, a local interpretability method based on
decision rules introduced by Marco Tulio Ribeiro, the author of
Lime.
I discovered for tabular data that the discretization method employed by LIME and Anchors impact their fidelity of the black box model and
I proposed a better discretization method to improve these methods.
Furthermore, I extended the latent research space used by Anchors to generate textual explanation by incorporating
pertinent negatives.
This internship ended with the publication of our work in the international conference
CIKM 2020.
2019
In charge of communication for the association of resident of Sherbrooke University. Agrus
I was in charge of the communication of an association with more than 500 adherent members. I created and organized events
through spreading them on social networks and communication with responsible.
I also took part to the organization of some events such as visiting a sugar shack or the winter carnaval in Québec.
2018
4 months internship in Lacodam team supervised by Luis Galárraga
I complete my university degree with an internship in a research laboratory. This was my first foot in the research, and I never leave it after.
This internship ended with the writing of an article concerning the mining of referring expressions published in the international conference
EDBT 2020..
During this internship I coded a programm called REMI, supervised by Luis Galárraga.
A referring expression (RE) is a description that identifies a set of instances unambiguously.
Mining REs from data finds applications in natural language generation, algorithmic journalism, and data maintenance.
Since there may exist multiple REs for a given set of entities, it is common to focus on the most intuitive ones, i.e., the most concise and informative.
REMI is a system that can mine intuitive REs on large RDF knowledge bases.
Vacataire Champs libres
During two years, I work as a student job in the library of Rennes Metropole, where is also located the Museum of Brittany, the Sciences Space and a conference hall.
I did public service, organized and sorted the books. Finally, I was in charge of the reception of the visitors.
Teaching
Introduction to Neural Network - M2 Students - at ENSAI. 2021-2022
Teaching of practical work and directed work of introduction to neural networks for last year students at ENSAI.
We introduce to students basic neural networks for tabular and image data on respectively
wine
and
Cifar10 datasets. The final project was to implement a
neural network to classify cats from dogs.
Oral Presentation for Technology Monitoring - M2 Students - at ISTIC, Rennes 1 University. 2020-2022
Evaluation of a report and an oral presentation of a scientific topic to be further investigated and reported.
The purpose is to challenge the student to embrace scientific information and technology as well as to learn how to synthesise
this information and present it in English.
(short term) prediction methods - M1 Students - at ISTIC, Rennes 1 University. 2020-2022
Teaching of practical work and directed work of prediction methods of time series for M1 students at the University of Rennes 1.
Throughout this teaching, I introduce prediction methods for time series such as exponential smoothing and auto regressive models.
The end of the year is completed by a
kaggle project I introduced on prediction of house prices.
Office tools for the statistician - L3 Students - at ENSAI. 2020-2022
Teaching of directed/practical work of office tools for the statistician such as Excel and Latex for two groups of first year student at ENSAI
(École nationale de la statistique et de l'analyse de l'information). This first teaching experience was a good introduction for the exercise.
Moreover, it was my only classroom teaching of the year due to the pandemic situation.
Object programming in Java - L2 Students - at ISTIC, Rennes 1 University. 2020-2022
Teaching of practical work in object programming for L2 student at the University of Rennes 1 in Java.
The purpose of this course is to discover the concept of object programming, notion of inheritance and collections such as trees and lists.
Throughout this teaching I learnt to teach distance learning course and interact with students through Microsoft Teams.
I implemented the final project which consisted in creating a kind of tower defense video game.
Publication
Julien Delaunay,
Luis Galárraga,
Christine largouët When Should We Use Linear Explanations?
Full paper at the Conference on Information and Knowledge Management (
CIKM 2022), Atlanta.
[Code]
Romaric Gaudel,
Luis Galárraga,
Julien Delaunay,
Laurence Rozé, Vaishnavi Bhargava.
s-LIME: Reconciling Locality and Fidelity in Linear Explanations.
Intelligent Data Analysis (
IDA 2022), Rennes.
[Preprint]
Julien Delaunay,
Luis Galárraga,
Christine largouët Improving Anchor-based Explanations.
Poster at the Conference on Information and Knowledge Management (
CIKM 2020), Galway.
[Preprint] [Presentation]
[Code]
Luis Galárraga,
Julien Delaunay,
Jean-Louis Dessalles.
REMI: Mining Intuitive Referring Expressions. International Conference on Extending Database Technology
(
EDBT/ICDT 2020), Copenhagen.
[Technical report]
[Full text]
[Presentation] [Code]
Personnal