ARC Algorithms and Representations for Continual Learning
- Coordinator
- Ecole Polytechnique Federale De Lausanne, Switzerland
- Runs
- 1 January 2027 to 31 December 2031
- Status
- Running
- EU funding
- €3.3m
- Total cost
- —
- Partners
- 1 organisations in 1 countries
- Type
- European Research Council grant
- Grant agreement
- 101267868
What the project is doing
In the project's own words, from its summary on CORDIS.
Massive data poses a fundamental challenge to machine learning (ML) since algorithm running times scalewith data size. As data grows faster than computational power, future problems can take longer to solvewithout economizing training via continual learning (CL). However, CL can result in performance drops inprior tasks, known as catastrophic forgetting (CF), a persistent issue in ML, including fine-tuning foundationmodels (FMs), blocking their full potential.
Contrary to the dogma that increasing more resources are critical for future FMs, our recent researchshows that computational investments can be recovered near-optimally without CF, hinting at a sustainable,future-proof ML paradigm. Our observation calls for new formulations in robust representation learningthat prevent catastrophic overfitting (CO) and new game-theoretic algorithms that avoid limit cycles (LC). However, understanding the trade-offs between data and computation in CL is at its infancy, especially fornon-linear models.
Existing methods are case-specific and do not rigorously offer sustainability, robustness, ortrustworthiness. The proposed ERC project ARC addresses these challenges by establishing mathematical foundations tomanage continuous data growth without CF or CO or LC and to optimize dynamic representations withnew robust and reinforcement learning formulations. We build new adaptive algorithms as well as decisionverification techniques that optimally handle increasing data sizes as well as dynamic models while matchingtheoretical runtime bounds.
We demonstrate our rigorous theory and robust learning frameworks in real-worldscenarios, such as neuro-rehabilitation and circuit design, to validate methods and explore trade-offs acrossdomains. Our approach views learning problems as dynamic and interactive, potentially revolutionizing ML by creatingholistic, sustainable systems with optimal performance across diverse applications under limited resources.
Topics
Who is involved
| Organisation | City | Country | Role |
|---|---|---|---|
| Ecole Polytechnique Federale De Lausanne | Lausanne | Switzerland | Coordinator |
Source: CORDIS, European Commission. Data of 5 Nov 2026, reused under the Commission's reuse policy. View on CORDIS · Grant DOI