Recerco l'aprenentatge continu a les xarxes neurals. Seeking to carve machines' memories in their hearts. PhD researcher at @tuytelaarslab
Leuven / Barcelona, Catalunya
Joined November 2016
- Tweets306
- Following338
- Followers66
- Likes1.2K
Pinned Tweet
Catastrophic forgetting: your model "forgets" a task, and probing confirms the info is gone. Case closed... right?
Not so fast
We found that sometimes tests like probing don't tell the full story. Features are still there, just misaligned or too crowded.
Let's dive in!🧵
Gotta act faster when I get an idea or I see it around soon after 😫
Nice work
Thrilled to introduce Mirror, our first "learning-from-experience" project! Assume you had a video generative model that, no matter your situation, could always generate video of what it would look like to solve a task. How can you put that plan into action? Clearly, you need to learn some video-to-action mapping - but where do you get the data for that? It turns out you can collect that data in an amazingly sample-efficient way by *incrementally following the video planner*! (1/n)
Sergi Masip retweeted
🇨🇳 | Una novia en China sorprendió a su futuro esposo durante la boda con una tarjeta gráfica ASUS RTX 5090 valorada en unos 12.000 dólares. El regalo fue tan inesperado que él terminó llorando de emoción.
When you can't get over 40% SR on a benchmark and suddenly you destroy it after one simple change to your method
Sergi Masip retweeted
Today, our colleague @ti_hess succeeded in his preliminary defence 🎉🎉🎉
Some members of the jury also shared their research with us!
- Joost van de Weijer (@CVC_UAB): post-hoc interventions for continual learning
- Andrea Soltoggio @asoltoggio: neuromodulation in lifelong RL
Sergi Masip retweeted
Last week I attended IEEE EAIS 2026 in Pisa to present “Lifelong Representations”, our survey on Continual SSL.
A big thanks to both the organizers for this opportunity, and to all of our co-authors for making this collaboration possible.
Link to the paper in the comments
Sergi Masip retweeted
#CoLLAs2026 was a very productive conference for our lab!
@ti_hess @JhaAbsk presented a poster on co-observation, a third facet of #continual_learning
@Atenrev presented:
- Putting a Face to Forgetting: cl meets mech interp
- Lifelong Representations: CSSL survey
Congrats! ❤️
Getting ready for day 3 of #CoLLAs2026! Come see our poster on Continual Self Supervised Learning to find out about promising directions in the field!
Tomorrow we present our poster on Putting a Face to Forgetting at #CoLLAs2026! If you're interested in understanding forgetting from a mechanistic perspective, drop by our poster! :D
I think I will start opening commissions to researchers who want a nice poster for their next conference (this one is still in progress)
Your vision model forgets. But did you know self-supervised ones forget considerably less, and too many people ignore this fact?
We surveyed the field of Continual Self-Supervised Learning (CSSL) for vision. Accepted at IEEE EAIS 2026 & presented at @CoLLAs_Conf next week!
🧵👇
We organize existing methods into 6 families based on how they fight forgetting:
🔵 Distillation (projected >> naive)
🟢 Weight regularization
🟠 Replay
🟣 Architectural methods
🟡 Model merging
🔴 Objective-level adaptation ← the most underexplored, IMO the most exciting
Big picture: CSSL needs to stop being "supervised CL but without labels". SSL has its own geometry, its own dynamics, and its own failure modes.
We need benchmarks that reflect that. And we need to test at scale.
Full survey 👉 arxiv.org/abs/2607.09785
#CoLLAs2026
