From discrete boosting iterations to differential equations in function spaces — where machine learning meets dynamical systems and operator theory.
Boosting iterates on a model problem approach the limiting flow F(t) = 1 − e−t as the learning rate ν vanishes.
This workshop introduces gradient boosting from both algorithmic and continuous-time perspectives. Particular emphasis will be placed on small and vanishing learning-rate regimes, where discrete boosting dynamics can be described through differential equations in function spaces.
The workshop will also highlight connections with dynamical systems, operator-based viewpoints and scientific machine learning.
Each round adds a weak learner hm, fitted to the negative gradient of the loss, scaled by the learning rate ν.
In the vanishing learning-rate limit, the boosting path follows an evolution equation in a space of functions.
Gradient boosting as an iterative procedure: weak learners, loss functions and the stagewise construction of a predictor.
How boosting behaves as the learning rate shrinks, and how its discrete dynamics are described through differential equations in function spaces.
Reading learning as an evolution in function space, and the operator-based viewpoints this opens up.
Where these continuous-time ideas meet scientific machine learning and the modelling of physical systems.
A mathematician working at the crossroads of statistics, probability and time-series analysis.
Youssef Esstafa is a faculty member and researcher in mathematics at Le Mans University, within the Department of Mathematics and the Laboratoire Manceau de Mathématiques. His research focuses on statistics, probability, stochastic processes, statistical inference and time-series analysis, with particular interest in long-memory models and FARIMA processes.
He has authored and co-authored numerous publications on statistical estimation, diffusion processes, nonparametric statistics, stochastic modelling and quantitative limit theorems. Alongside his research, he teaches time-series analysis and data science, and contributes to work at the intersection of mathematical statistics, probability and modern data analysis.
Long-memory time-series models with dependent innovations
During ICAOTA'2027, March 22–24, 2027. The session slot will appear in the conference programme.
Steigenberger Marhaba Thalasso Hammamet
Hammamet, Tunisia
Researchers and students in operator theory, differential equations, dynamical systems and machine learning.
Open to ICAOTA'2027 participants.
Register to attend
Take part in ICAOTA'2027 and join the workshop alongside the conference's lectures and sessions.
Register to attend