A new research paper explores catastrophic forgetting in Walrus-1.3B, a physics foundation model. The study found that adapting the model to new data, specifically simulated flow around a cylinder and then experimental data of the same system, significantly degraded its performance on original pretraining tasks. Errors increased up to 35 times after just 20 epochs of finetuning on simulations. The research suggests that replaying a small percentage of the original simulated data during adaptation to real-world data can recover most of the lost performance while still allowing effective adaptation to the new experimental dynamics. AI
IMPACT Highlights potential limitations in adapting scientific foundation models to new data, suggesting replay strategies as a mitigation for catastrophic forgetting.
RANK_REASON This is a research paper detailing findings about a specific model's behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
