Maxime Fuccellaro's PhD Thesis

25 November 2024

Most recent news

(list of all news)

Round Table Women in Tech

(05 December 2024)

At the initiative of Betclic, as part of their "Diversity Week", I was able to participate in a round table (without a table) with Amélie Benoit (Women in Tech Bordeaux)...


Maxime Fuccellaro's PhD Thesis

(25 November 2024)

Accompanying a student to the achievement of a Phd is always an adventure. With Maxime Fuccellaro, whom I co-supervised with Akka Zemmari, it was an opportunity to discover a new...


Feedback on the LLM Afternoon by the Trustworthy AI Chair

(20 November 2024)

Last Wednesday, November 20th, as part of the activities of the Trustworthy AI Chair, we offered two hours covering the main themes related to LLMs, at ENSEIRB-MATMECA. You can find...


Jury at HackRobot Hackathon

(17 November 2024)

A Sunday afternoon watching robots and innovations everywhere? I had the pleasure of being invited to be a jury member at the Hack1Robo organized by INRIA Bordeaux South West. From...


Alexis Juven's PhD Thesis

(12 November 2024)

One of the ways to realize how AI permeates all areas of society is to measure the extent of its applications. I was happy to be the reviewer of Alexis...

Accompanying a student to the achievement of a Phd is always an adventure. With Maxime Fuccellaro, whom I co-supervised with Akka Zemmari, it was an opportunity to discover a new field: model drift in machine learning. Driven by the issues developed in the Trustworthy AI chair, I found what makes research exciting, that is, a new field where the state of the art is still incomplete.

The classic approach in machine learning is to train a supervised model that achieves the best accuracy compared to the true classification of the target document. However, once in production, the input data describing the document can slowly drift, change, or become noisy. This phenomenon is well known and leads to increasingly poor model classifications.

However, few current works are realistic in addressing this problem. Many assume the ability to measure ground truth and thus the loss of accuracy. But what is the point of having a predictive model if ground truth is available?

In this PhD work, Maxime proposed methods that do not assume ground truth. Additionally, he focused on virtual drift, which is an observed drift in input data that does not require model retraining. This can have a significant impact on the energy consumption of models in production, as less model retraining will be advised by the system.

It was a pleasure to see this work come to fruition, with an eclectic jury that we had the pleasure of hosting in Bordeaux. Thanks to all:

Group photo of the thesis

Thanks to Mangrove for funding this CIFRE thesis, which allowed a significant focus on fundamental research while offering Maxime great freedom.

Thanks (and congratulations) to Maxime for successfully completing this thesis and maintaining an unwavering passion for “creating solutions,” developing new tools, seeking to measure them, following his own ideas, and integrating the diagrams drawn on the board into real classifiers.

PS: As soon as the final version of the manuscript is available, I will update this post.

(Last Modified date: 28 November 2024)
(version française)
(list of all news)