
Fresh from the press: "Applied Stochastic Differential Equations" with @simosarkka and published by @CambridgeUP. Physical books: cambridge.org/fi/academic/su… Online PDF version: users.aalto.fi/~asolin/sde-b… Codes: github.com/AaltoML/SDE

Associate Professor in Machine Learning @AaltoUniversity. Enjoying statistical machine learning. @ELLISforEurope Scholar. @yaf_fi

Fresh from the press: "Applied Stochastic Differential Equations" with @simosarkka and published by @CambridgeUP. Physical books: cambridge.org/fi/academic/su… Online PDF version: users.aalto.fi/~asolin/sde-b… Codes: github.com/AaltoML/SDE

6/ Joint work with @bhrobson_, @SanteriMentu, and @wszhao_robot at ELLIS Institute Finland and Aalto University. Check out our paper, project page with interactive examples, and code: 📄 Preprint: arxiv.org/abs/2610.06226 💻 Project page: aaltoml.github.io/LeAVJEPA

5/ In the video, audio-to-video attention follows the sound-producing objects, although the model received no localisation supervision. With attentive probes on frozen ViT-L features, we obtain 36.0 mAP on AudioSet-20K and 91.3% accuracy on ESC-50.

4/ Embedding alignment brings the views together and LeJEPA's SIGReg regularizer prevents collapse. Pretraining needs no EMA teacher, separate predictor, reconstruction decoder, or contrastive negatives.

3/ The important detail is how we construct the views. When every view contains both modalities, matching them can be solved by relying on just one. Modality dropout makes that shortcut harder, and our controlled ablation shows a clear improvement in the learned features.

2/ In #LeAVJEPA, we extend #LeJEPA (by @randall_balestr and @ylecun) to audio and video using one shared transformer. During training, 🎤 audio-only and 📹 video-only views learn to match a representation formed from both modalities.

🔥 #LeAVJEPA: A simple recipe for self-supervised learning from sound and vision 🔥 One shared encoder, trained without class labels. Dropping a modality makes audio & video learn a common representation. 🧵

Submissions will open soon for #AISTATS 2027 More info (soon) at: virtual.aistats.org/Conferen… Submission opens Sept 15th 2027 at 🌍 🕜 - Beijing: 7:30 PM - Paris: 1:30 PM - Montreal/NY: 7:30 AM - San Francisco: 4:30 AM - AOE: 11:30PM (Sept. 14th)

#AISTATS 2027 Call for Paper is out! This year's edition is taking place in Montreal on May 3-6th, 2027 🇨🇦 Aymeric Dieuleveut and I are chairing virtual.aistats.org/Conferen… New this year: AI review, and more! Abstract deadline: September 29, 2026 Paper deadline: October 6, 2026

I had the privilege of sitting down with @kchonyc for our new AI podcast "And humanity created intelligence". we discussed the backstory behind the attention paper as well as... 🧵

I joined Aalto University's podcast "And humanity created intelligence" for an episode on AI safety, including the recent OpenAI / Hugging Face attack. Many thanks to @ericmalmi and @arnosolin for a great conversation! youtube.com/watch?v=nx0CeYeL…

Links to the podcast: 🎧 Spotify: open.spotify.com/show/033LOU… 🎧 Apple Podcasts: podcasts.apple.com/us/podcas… 🎧 YouTube: youtube.com/watch?v=nx0CeYeL…

We also talk about the early days of OpenAI, RLHF, and the ideas behind ChatGPT, immediate vs. catastrophic AI risks, AI consciousness and model welfare -- and, of course, p(doom). 4/n

The discussion around OpenAI's agents escaping their sandbox is still very much ongoing and gets right to the heart of something we discuss in the episode: when does AI safety stop being about hypothetical future scenarios and become an issue we are dealing with today? 3/n

Our guest is Katarina Slama, a former #OpenAI researcher who worked on language-model alignment and co-authored the influential #InstructGPT paper. Today, she works independently on AI safety. 2/n

🎙️Katarina Slama, ex-OpenAI, discusses AI alignment and risks in our new episode of "And Humanity Created Intelligence" -- and, for the first time, this one is in English. 1/n