Infinite Schedules and the Benefits of Lookahead
Published:
TL;DR: Knowing the next training checkpoint in advance (“lookahead”) helps to set the learning rate. In the limit, the classical square-root schedule appears on the horizon.
Blog posts
2025
2024
A Bibliography Database for Machine Learning
Published:
Getting the correct bibtex entry for a conference paper (e.g. published at NeurIPS, ICML, ICLR) is annoyingly hard: if you search for the title, you will often find a link to arxiv or to the pdf file, but not to the conference website that contains the bibtex.
How to jointly tune learning rate and weight decay for AdamW
Published:
TL;DR: AdamW is often considered a method that decouples weight decay and learning rate. In this blog post, we show that this is not true for the specific way AdamW is implemented in Pytorch. We also show how to adapt the tuning strategy in order to fix this: when doubling the learning rate, the weight decay should be halved.
2023
Optimization Nuggets: Stochastic Polyak Step-size, Part 2
Published:
Fabian Pedregosa invited me to write a joint blog post on a convergence proof for the stochastic Polyak step size (SPS).
Decay No More
Published:
I wrote a blog post which got published at the ICLR blog post track 2023. The post is titled Decay No More and explains the details of AdamW and its weight decay mechanism. Check it out here.
2022
Solve it all and solve it fast: using numba for optimization in Python
Published:
When implementing optimization algorithms, we typically have to balance the following goals:
A collection of resources for creating open-source software packages
Published:
Making your research code open-source, tested and documented is quite simple nowadays. This post gives an overview of the most important steps and collects useful ressources, e.g. tutorials for Readthedocs, Sphinx (Gallery) and unit testing in Python.