Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Research
arXiv CS.AI·1d ago·5 min read
Relational Semantic Grounding for Test-Time Discovery
TL;DR
SEGA, a new framework for Test-Time Discovery (TTD), enhances the interpretability of discovered classes by using large language models to generate attribute-rich descriptions. This approach transforms anonymous clusters into meaningful semantic concepts, improving TTD performance in open-world scenarios.
✦ Why It Matters
Implement SEGA to improve the interpretability of your TTD systems today.
Key Takeaways
- Explore SEGA for enhanced class interpretability in TTD.
- Utilize large language models for generating attribute descriptions.
- Implement RSG to improve candidate concept selection.
- Focus on semantic alignment for better open-world performance.
How It Works
SEGA constructs an attribute-anchored visual-semantic space during training, allowing it to ground unknown samples by comparing them to known-class anchors. At test time, it prompts a large language model to generate candidate concepts based on these comparisons, which are then reranked to select the most relevant concept.
This process transforms anonymous clusters into interpretable semantic concepts, enhancing the model's ability to recognize and label new classes.