iiinova — where your digital self keeps evolving

Thomas Michel

@thomas-michel.bsky.social

PhD student @ Inria Scool team. Working on responsible AI, differential privacy, robustness & sequential decision-making 🧠

Posts Replies Media

Thomas Michel @thomas-michel.bsky.social · 02/05/2026

Big thanks to my advisors Debabrota Basu and Emilie Kaufmann. 📄 arxiv.org/abs/2508.06377

arxiv.org

DP-SPRT: Differentially Private Sequential Probability Ratio Tests

We revisit Wald's celebrated Sequential Probability Ratio Test for sequential tests of two simple hypotheses, under privacy constraints. We propose DP-SPRT, a wrapper that can be calibrated to achieve...

020

Thomas Michel @thomas-michel.bsky.social · 02/05/2026

We also prove a lower bound for any ε-DP sequential test, revealing two regimes: - one where privacy is essentially free - one where it dominates the sample complexity DP-SPRT with Laplace noise matches this lower bound in the small-error, close-hypotheses regime.

120

Thomas Michel @thomas-michel.bsky.social · 02/05/2026

A key building block: OutsideInterval, a DP mechanism that monitors whether a stream of queries leaves a *two-sided* range. Tools like AboveThreshold are one-sided, using them twice burns privacy budget. OutsideInterval handles both bounds at once and is reusable beyond SPRT.

110

Thomas Michel @thomas-michel.bsky.social · 02/05/2026

Our main contribution: DP-SPRT, a private SPRT supporting several noise mechanisms (Laplace for ε-DP, Gaussian for Rényi DP). Thresholds are calibrated theoretically, so type I/II error control holds out of the box rather than through Monte Carlo tuning.

100

Thomas Michel @thomas-michel.bsky.social · 02/05/2026

Sequential hypothesis testing lets you stop early once the evidence is conclusive. The catch on sensitive data: the *stopping time itself* leaks information. If a clinical trial halts and returns a result, you can often deduce things about the enrolled patients.

100

Thomas Michel @thomas-michel.bsky.social · 02/05/2026

AISTATS 2026 starts this week in Tangier 🇲🇦 I'll be presenting our paper "DP-SPRT: Differentially Private Sequential Probability Ratio Tests" 🎤 Spotlight talk : Monday, 11:30 🪧 Poster : Monday, 15:00 A short thread 🧵👇

Illustration of the principle behind DP-SPRT algorithm

120

Thomas Michel @thomas-michel.bsky.social · 24/03/2026

Thank you !

000

Reposted by Thomas Michel

EWRL @ewrl-org.bsky.social · 27/02/2026

Announcing the 19th European Workshop on Reinforcement Learning! We are looking forward to seeing you in Lille, France on October 5th — 7th, 2026. More details to come (ewrl-org.github.io/ewrl-2026/in...) 👀

ewrl-org.github.io

EWRL 2026

EWRL 2026 is the 19th European Workshop on Reinforcement Learning, held October 5–7 in Lille, France.

043

Thomas Michel @thomas-michel.bsky.social · 18/03/2026

DP-SPRT accepted at AISTATS 2026! 🎉 How do you decide when to stop a sequential test without leaking private information? We propose a near-optimal differentially private sequential test, with a reusable private monitoring mechanism. With Debabrota Basu & Emilie Kaufmann. arxiv.org/abs/2508.06377

arxiv.org

DP-SPRT: Differentially Private Sequential Probability Ratio Tests

020

Thomas Michel @thomas-michel.bsky.social · 20/02/2026

We turned 1970s arcade games into GPU-accelerated RL environments. Accepted at ICLR 2026! 🎉 OCTAX emulates CHIP-8 games in JAX. 20+ games, 14× faster than EnvPool, fully vectorized. Retro games, modern hardware. 🕹️ w/ Waris Radji & Hector Piteau. Let us know what you think! warisradji.com/octax-page/

130

Reposted by Thomas Michel

Clément Moulin-Frier @clemmoulinfrier.bsky.social · 08/10/2025

Exploring Flow-Lenia Universes with a Curiosity-driven AI Scientist: Discovering Diverse Ecosystem Dynamics developmentalsystems.org/Exploring-Fl... Authors: @thomas-michel.bsky.social and @mcvjetko.bsky.social (both attending), @hamongautier.bsky.social and @pyoudeyer.bsky.social

developmentalsystems.org

Exploring Flow-Lenia Universes with Curiosity-driven AI

Automated discovery of diverse ecosystem dynamics in Flow-Lenia using AI-driven exploration. Features interactive visualization of 2000+ discovered evolutionary patterns.

111