✨Thrilled to see EurIPS launch — the first officially endorsed European NeurIPS presentation venue!
👀 But NeurIPS now requires at least one author to attend in San Diego or Mexico (and not just virtually as before). This is detrimental to many. Why not allow presenting at EurIPS or online?
1/4
Boys and girls receive similar maths scores at the start of school, but boys pull ahead of girls after just four months
https://go.nature.com/4kAk1Kz
This work opens the way to more secure distributed systems, for collaborative machine learning, and beyond!
We show that it can be combined with many robust averaging options and the resulting algorithm is highly resilient: the performance remains strong when the number of misbehaving computers doesn’t exceed a threshold. For an averaging technique we introduce, this threshold is optimal up to a factor 2!
We investigate this challenge in a setting where computers communicate in peer-to-peer, without relying on a central server.
Our approach builds on a simple idea: each computer can treat its information as a trustworthy reference and consider messages that differ too much as suspicious.
Can many computers train a model together, when some send wrong information? Is it harder when computers communicate with only a small number of the other ones?
arxiv.org/abs/2410.10418
Collaborative work with Aymeric Dieuleveut & Hadrien Hendrikx to be presented at #ICML2025.
