Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·10d ago·5 min read
Useful to Whom? Sample Value Is Defined Only Relative to the Learner
TL;DR
Coreset selection reveals that the value of training samples is relative to the learner's architecture. Adjusting model parameters, like width, can significantly shift the optimal sample selection boundary.
✦ Why It Matters
Evaluate your data selection strategy based on the specific architecture of your model to optimize training.
Key Takeaways
- Adjust model width to optimize sample selection for training.
- Consider input grid size when selecting training samples.
- Evaluate sample selection strategies based on the learner's architecture.
- Use Vision Transformers for improved coverage in sample selection.
How It Works
The study demonstrates that the effectiveness of training samples is not fixed but varies with the model's architecture. By manipulating the width of ResNet-18, researchers observed a significant shift in the optimal number of samples needed, indicating that different models value the same data differently.
Additionally, changing the input grid size and stride settings allowed for further adjustments to the sample selection boundary.