Catastrophic Forgetting Persists at Frontier Scale
A. Lindqvist, M. Ferreira, J. Osei — Independent
Abstract
Several methods for continual learning without catastrophic forgetting report success at research scale. We replicate five such methods at frontier scale and find that forgetting reappears in all cases. We characterise the scale at which each method fails and argue that the mechanisms proposed to explain their success do not survive the transition.
Key findings
- All five methods show forgetting reappearing at frontier scale
- Failure scale differs by method but the pattern is consistent
- Proposed mechanisms do not survive the scale transition
- Retrieval-based scaffolding continues to work and is not continual learning
Limitations
Published at equal prominence to the findings. A paper’s limitations are usually the part that determines how much its result should move your beliefs.
- Replication limited to five methods
- Compute cost prevented exhaustive hyperparameter search
- Negative results are hard to distinguish from insufficient tuning