Projects
stable-worldmodel
stable-worldmodel (SWM) is an open-source platform for standardized and
reproducible world-modeling research and evaluation. It provides a high-performance data layer,
clean and well-tested implementations of modern world-model baselines and planning solvers, and a
broad suite of environments and tasks with controllable visual, geometric, and physical factors of
variation — so dynamics understanding, control, representation quality, and out-of-distribution
generalization can be evaluated systematically under one scalable framework.
DTE — Dynamic Training Engine
A generic, block-based engine for training neural networks with adaptive and recursive execution, enabling composable training strategies without hard-coding specific algorithms. It ships a Tiny Recursive Model (TRM) baseline as a reference and supports multi-GPU training via PyTorch Distributed Data Parallel (Python 3.10+, Linux).
Regularized Nonlinear Acceleration (RNA)
Reference implementations — in both MATLAB and Python — of the Regularized Nonlinear Acceleration algorithm from my work with d'Aspremont and Bach. RNA is a generic scheme that speeds up optimization methods (e.g. gradient descent) by computing a nonlinear average, an extrapolation, of their iterates via a small linear system.
Book on Acceleration in Optimization
This monograph covers some recent advances in a range of acceleration techniques frequently used in convex optimization. We first use quadratic optimization problems to introduce two key families of methods, namely momentum and nested optimization schemes. They coincide in the quadratic case to form the Chebyshev method.
We discuss momentum methods in detail, starting with the seminal work of Nesterov, and structure convergence proofs using a few master templates, such as that for optimized gradient methods. We further cover proximal acceleration, at the heart of the Catalyst and Accelerated Hybrid Proximal Extragradient frameworks. We conclude by discussing restart schemes.
Non-exhaustive list of methods covered:
- Chebyshev's method and Polyak's momentum (a.k.a. Heavy-ball method)
- (Regularized) Nonlinear Acceleration
- Nesterov's scheme and Optimized Gradient Descent
- Catalyst scheme and Güler / Monteiro-Svaiter acceleration
- Restart schemes
FAST Toolbox for Matlab
FAST (Finally An SDDP Toolbox) is an easy-to-use Stochastic Dual Dynamic Programming (SDDP) toolbox for Matlab. It helps you model and solve your problem easily and quickly. The goal of SDDP is to solve large-scale stochastic problems with an algorithm based on the cutting-plane method.
- Energy production — accurate solutions when dealing with uncertain energy supply or demand.
- Factory production line — plan production to maximize revenue under unknown demand.
- Financial products — deal with random portfolios in the stock exchange business.
JSR Toolbox for Matlab
I partially contributed to the JSR toolbox by implementing a few algorithms. The Joint Spectral Radius of a set of matrices characterizes the maximal asymptotic rate of growth of a product of matrices taken in this set when the length of the product increases. It is known to be very hard to compute, and in recent years many different methods have been proposed to approximate it.
These methods have different advantages, depending on the application, the type of matrices, the desired accuracy or running time, etc. This toolbox aims to provide the practitioner with the best available methods and an easy tool for researchers to compare the different algorithms.