Distribution shift in ecological data: generalization vs. specialization,
summary
Ecological vision faces concrete tradeoffs: global models improve aggregate accuracy but fail local decisions, while specialized local models or fine-tuning often outperform them for actionable tasks like salmon escapement counting or tree canopy carbon estimates. Experiments show internet-scale data still misses true robustness (LAION filtering reveals gaps), labeled target data remains most valuable, and many domain-adaptation gains vanish under fair baselines that include EMA, strong augmentation and better initialization. Published unified code for object-detection adaptation shows feature alignment 2.6% and distillation 4.6% gains, batch composition can shift results ~15%, and subset selection plus compression yields practical, runnable specialized models.
AI-written summary · about
chapters (11)
- 0:00Ecological data challenges
- 2:08Why ecological data is hard
- 6:50Distribution shifts
- 10:06Foundation models myths
- 15:48Adapting models to shifts
- 18:57Domain adaptation failures
- 25:50Robust domain adaptation
- 28:40Specializing models
- 31:12Deployment-specific models
- 35:08Global vs local models
- 37:37Q&A
related talks
31mIndividual Gaps in Prediction Generalization
01h 01mDistribution Shift as Underspecification, and What We Might Do About It
39mChanging Realities - The Final Frontier for Machine Learning?
53mAre We Measuring the Right Thing? Distribution Shift Lessons for Federated Learning
36mWhen the Only Way to Scale up, is to Scale Down
48mSetting the stage for AI for biodiversity
01h 10mModern paradigms of generalization, the heliocentric model of Aristarchus,...
53mGeneralization insights from actual cognition