Rahaf Aljundi

This fall, during a Dagstuhl seminar on continual learning, we discussed with various researchers from the field the roadmap for continual learning. We converged to one view: modular memory is the key to continual learning agents, as outlined in here arxiv.org/pdf/2603.01761

How do you give an RL agent useful long term memory when it needs to act over thousands of steps? Storing everything in-context is expensive, text summaries lose detail and plain recurrence struggles with long horizons. Our NeurIPS Spotlight paper explores a simple idea 🧵:

@CVPR @BinTrash @CVPR what is the procedure for submitting an appeal? I couldn't find anything.

@BinTrash Also oral/poster decisions have NOT been released.

What is the start and end date of CVPR including all workshops, expo, tutorials and main conference? Google says June 10-17 and CVPR website says June 11-15. Please RT and help me @CVPR #CVPR2025 #cvpr

If you are at @eccvconf and interested in a postdoc position on efficient multimodal models, please ping me for a chat.

If you are working on continual learning with foundation models and you have something cool to share, please consider submitting to our NeurIPS workshop sites.google.com/u/0/d/1VBppSNb…

Are you interested in efficient & continual fewshot updates for VLMs with zero replay and close to zero forgetting? Check our work on PEFT of VLMs 2407.16526 (arxiv.org)

Do you know that you can significantly improve VLMs performance such as LLaVA and MiniGPT2 by only updating the vision encoder separately and plug it back in the corresponding VLMs? 2407.16526 (arxiv.org)

@_hamed_hemati Yes retraining from scratch would be fun for such a model ;)

This is why we need continual learning ;)

We are looking for excellent students for research internships on Continual Learning at Toyota Motor Europe. Please share and contact me if you are interested. social.icims.com/viewjob/pt1669…

[Worskhop on Theory of Continual learning (ICML'21)] Thrilled to announce our awesome speaker line up: sites.google.com/view/cl-theory… You can ask questions to our speakers here: docs.google.com/forms/d/e/1FAI… Happening this Friday 23rd !🔥🚀

We [EXCEPTIONALLY] extend the deadline of our workshop on "Theory of Continual Learning" (ICML'21) to June 11th (AoE) more information at: sites.google.com/view/cl-theory…

+ 8min video youtube.com/watch?v=wfb9UV… + code! github.com/optimass/Maxim…

Our new continual learning survey with @matt_dl96 and @mmasana is now on arxiv. We compare 10 SOTA continual learning methods on 2 benchmarks. We investigate the effect of model capacity, regularization, task ordering and so much more. Check it out at: arxiv.org/abs/1909.08383

Most continual learning approaches assume that training on a new task will cause equal interference over previously learned tasks. With "Maximally Interfered Retrieval" we relax this assumption by automatically rehearsing on tasks currently undergoing the most forgetting.

Interested in Continual (Lifelong) Learning? Come to the Workshop on Multi-Task and Lifelong Reinforcement Learning tomorrow (Saturday) @ ICML for posters and oral on how to rehearse on on older tasks efficiently!

Our new paper on continual learning with no task boundaries

@mikarv @lilianedwards I guess telling the network not to forget important thing is different from asking it to forget something

Given a limited model capacity and unlimited evolving data, it is not possible to preserve all the previous knowledge that an agent has learned. How about learning what (not) to forget instead? That's what we humans do and our method aims at arxiv.org/abs/1711.09601

We propose a new solution to the catastrophic forgetting when learning tasks sequentially using task autoencoders. arxiv.org/pdf/1704.01920…