user avatar
Fabian Pedregosa
@fpedregosa
Keeping the gradients flowing since 2013. Loves open source. Sometime blogs and writes papers.
Geneva, Switzerland
fa.bianp.net
Joined December 2010
612
Following
6,082
Followers
Posts
  • Pinned
    user avatar
    Fabian Pedregosa
    @fpedregosa
    Jul 25, 2024
    Beyond proud to have been part of the team behind this milestone. I believe the combination of LEAN + RL + LLMs will change how we do maths forever. This is just the beginning!
    Some of the people behind this achievement
    Animation with a blue background and faint white outlines of a cube, sphere and mathematical symbols surrounding a central glowing sphere with lines criss crossing through it.
    user avatar
    Google DeepMind
    @GoogleDeepMind
    Jul 25, 2024
    We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈 It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 dpmd.ai/imo-silver
    33K
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Feb 14, 2022
    Are you a PhD student interested in optimization and/or machine learning for code? Come for a summer internship with us @GoogleAI! You'll be working with me and @dtarlow2 on problems related to optimization and code generation in practice!
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Dec 16, 2021
    I'm starting a series of blog posts on "optimization nuggets": short and beautiful proofs in optimization (let me know what you think!). First one shows that SGD converges exponentially fast to a neighborhood of the solution:
    fa.bianp.net
    Optimization Nuggets: Exponential Convergence of SGD
    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...
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Mar 2, 2021
    Third part of my blog series on optimization and polynomials: "A hitchhiker guide to momentum". fa.bianp.net/blog/2021/hitc… Using properties of Chebyshev polynomials, we can categorize the convergence of momentum into three regions.
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Aug 29, 2022
    After a 4-year hiatus, I'm bringing back Frank-Wolfe! An overlooked aspect of this algorithm is the step-size selection. In this post I discuss different step-size strategies, and how to construct a backtracking line-search strategy that works:
    fa.bianp.net
    Notes on the Frank-Wolfe Algorithm, Part III: backtracking line-search
    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...
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Apr 8, 2020
    I'm back into blogging with a post the link between optimization methods and polynomials: fa.bianp.net/blog/2020/poly…
    GIF
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Jun 14, 2023
    🚨 New blog post alert🚨: On the Convergence of the Unadjusted Langevin Algorithm fa.bianp.net/blog/2023/ulaq/ The Langevin algorithm is a marvel of optimization: it allows to solve an infinite-dimensional problem (approximating a probability distributions) in finite time.
    fa.bianp.net
    On the Convergence of the Unadjusted Langevin Algorithm
    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...
    72K
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Apr 14, 2021
    New blog post: Acceleration without Momentum. After two blog posts on momentum, now one on how to get the same effect without it, just through some well-chosen step-sizes (🤯). fa.bianp.net/blog/2021/no-m…
    Young's method has the same last-iterate convergence rate than the Chebyshev iterative method on quadratic problems. Rate on previous iterates is substantially different and depends on the step-size ordering.
    Young's method. A funny-looking step-size, but no momentum.
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Oct 6, 2025
    Life update: The past 2 years on AI for math have been a wild ride. In that time, our team competed in the International Math Olympiad, winning silver 🥈 and then gold 🥇. In my next chapter I'll be helping make Gemini the best AI model for code. Excited on what's next
    29K
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Jul 15, 2021
    Everything you wanted to know about Random Matrix Theory (but were afraid to ask): join us Monday at the @icmlconf tutorial on #RMT and #ML. Materials coming together at random-matrix-learning.github.io
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Aug 24, 2023
    Today's office view. Can't complain.
    25K
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Jul 19, 2021
    Join us tomorrow for the #ICML2021 tutorial on random matrix theory and machine learning (random-matrix-theory.github.io). This tutorial has 4 parts:
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    May 30, 2022
    Why do cyclical step-sizes work? And more importantly, when _don't_ they work? I wrote a blog post about it fa.bianp.net/blog/2022/cycl… based on a recent paper. Ready to follow me down the rabbit hole? 🐇
    GIF
  • user avatar
    Fabian Pedregosa
    @fpedregosa
    Oct 24, 2022
    I recently learned about the Russian Roulette (in statistics), a mind-boggling trick to compute an unbiased estimator of the limit. To understand it better I wrote a blog post about it:
    fa.bianp.net
    The Russian Roulette: An Unbiased Estimator of the Limit
    The idea for what was later called Monte Carlo method occurred to me when I was playing solitaire during my illness. Stanislaw Ulam, Adventures of a Mathematician The Russian Roulette offers a simple...