Announcing Native BM25 Ranking in AlloyDB and Cloud SQL
cloud.google.com·1d ago
AI Research
arXiv CS.AI·Jun 11·5 min read
Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders
TL;DR
Sparse autoencoders, a type of neural network, often produce unstable features due to their dependence on random initialization seeds. This study investigates the seed dependence in sparse autoencoders and introduces a method to identify reproducible subspaces.
✦ Why It Matters
Engineers can leverage reproducible subspaces to enhance the stability and reliability of sparse autoencoders in their projects.
Key Takeaways
- Evaluate feature stability in your sparse autoencoders.
- Identify and pool unique features across training seeds.
- Focus on stable features for better model performance.
- Explore low-rank subspace structures in your models.
How It Works
The study quantifies feature stability by calculating the probability of feature recurrence across different training seeds. It distinguishes between stable and unstable features, revealing that stable features are essential for effective model performance, while unstable features often reflect low-frequency patterns rather than meaningful representations.