This blog post discusses the convergence rate of the Stochastic Gradient Descent with Stochastic Polyak Step-size (SGD-SPS) algorithm for minimizing a finite sum objective. Building upon the proof of the previous post, we show that the convergence rate can be improved to O(1/t) under the additional assumption that …

The stochastic Polyak step-size (SPS) is a practical variant of the Polyak step-size for stochastic optimization. In this blog post, we'll discuss the algorithm and provide a simple analysis for convex objectives with bounded gradients.

The Langevin algorithm is a simple and powerful method to sample from a probability distribution. It's a key ingredient
of some machine learning methods such as diffusion models and differentially private
learning.
In this post, I'll derive a simple convergence analysis of this method in the special case when the …

Backtracking step-size strategies (also known as adaptive step-size or approximate line-search) that set the step-size based on a sufficient decrease condition are the standard way to set the step-size on gradient descent and quasi-Newton methods. However, these techniques are much less common for Frank-Wolfe-like algorithms. In this blog post I …

Six: All of this has happened before. Baltar: But the question remains, does all of this have to happen again? Six: This time I bet no. Baltar: You know, I've never known you to play the optimist. Why the change of heart? Six: Mathematics. Law of averages. Let a complex …

When an optimization problem has multiple global minima, different algorithms can find different solutions, a phenomenon often referred to as the implicit bias of optimization algorithms. In this post we'll characterize the implicit bias of gradient-based methods on a class of regression problems that includes linear least squares and Huber …

This is the first of a series of blog posts on short and beautiful proofs in optimization (let me know what you think in the comments!). For this first post in the series I'll show that stochastic gradient descent (SGD) converges exponentially fast to a neighborhood of the solution.

While the most common accelerated methods like Polyak and Nesterov incorporate a momentum term, a little known fact is that simple gradient descent –no momentum– can achieve the same rate
through only a well-chosen sequence of step-sizes. In this post we'll derive this method and through simulations discuss its practical …

I've seen things you people wouldn't believe.
Valleys sculpted by trigonometric functions.
Rates on fire off the shoulder of divergence.
Beams glitter in the dark near the Polyak gate.
All those landscapes will be lost in time, like tears in rain. Time to halt.

We can tighten the analysis of gradient descent with momentum through a cobination of Chebyshev polynomials of the first and second kind. Following this connection, we'll derive one of the most iconic methods in optimization: Polyak momentum.

There's a fascinating link between minimization of quadratic functions and polynomials. A link
that goes
deep and allows to phrase optimization problems in the language of polynomials and vice versa.
Using this connection, we can tap into centuries of research in the theory of polynomials and
shed new light on …

A naive implementation of the logistic regression loss can results in numerical indeterminacy even for moderate values. This post takes a closer look into the source of these instabilities and discusses more robust Python implementations.

This blog post extends the convergence theory from the first part of these notes on the
Frank-Wolfe (FW) algorithm with convergence guarantees on the primal-dual gap which generalize
and strengthen the convergence guarantees obtained in the first part.

Most proofs in optimization consist in using inequalities for a particular function class in some creative way.
This is a cheatsheet with inequalities that I use most often. It considers class of functions that are convex,
strongly convex and $L$-smooth.

The main contribution is to develop a parallel (fully asynchronous, no locks) variant of the SAGA algorighm. This is a stochastic variance-reduced method for general optimization, specially adapted for problems …

Announce: first public release of lightning!, a library for large-scale linear classification, regression and ranking in Python. The library was started a couple of years ago by Mathieu Blondel who also contributed the vast majority of source code. I joined recently its development and decided it was about time for …

Together with other scikit-learn developers we've created an umbrella organization for scikit-learn-related projects named scikit-learn-contrib. The idea is for this organization to host projects that are deemed too specific or too experimental to be included in the scikit-learn codebase but still offer an API which is compatible with scikit-learn and …

Recently I've implemented, together with Arnaud Rachez, the SAGA[^{1}] algorithm in the lightning machine learning library (which by the way, has been recently moved to the new scikit-learn-contrib project). The lightning library uses the same API as scikit-learn but is particularly adapted to online learning. As for the SAGA …

Cross-validation iterators in scikit-learn are simply generator objects, that is, Python objects that implement the __iter__ method and that for each call to this method return (or more precisely, yield) the indices or a boolean mask for the train and test set. Hence, implementing new cross-validation iterators that behave as …

Due to lack of time and interest, I'm no longer maintaining this project. Feel free to grab the sources from https://github.com/fabianp/nbgallery and fork the project.

TL;DR I created a gallery for IPython/Jupyter notebooks. Check it out :-)