Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·24d ago·5 min read
Can Data Attribution Filter Out Subliminal Learning? Not Reliably
TL;DR
Gradient-based data attribution methods can partially identify subliminal learning in language models, but their effectiveness varies significantly across different models and settings.
✦ Why It Matters
Consider implementing EK-FAC for data attribution to better manage subliminal learning in your models.
Key Takeaways
- Evaluate gradient-based attribution methods for your language models.
- Use EK-FAC for better identification of subliminal learning.
- Be aware of inconsistent performance across different model settings.
- Consider token-level filtering over sample-level for improved results.
How It Works
Gradient-based attribution methods analyze the gradients of model outputs with respect to input data to determine which training examples most influence specific behaviors. By focusing on the gradients, these methods can identify data points that contribute to subliminal learning, even when the data lacks semantic relevance.