Researchers have characterized the cost of achieving valid per-class coverage in recognition systems when distribution shift occurs between training and testing data. They found that while split conformal prediction maintains marginal coverage, per-class coverage can fail significantly. The study proposes methods to restore per-class validity, noting that a small number of target labels per class are sufficient, but achieving efficiency alongside validity requires a more substantial label count that scales with the number of classes and tolerance for efficiency loss. A case study using skeleton action recognition demonstrated that per-class calibration with source labels alone can recover a significant portion of the coverage gap. AI
IMPACT Provides theoretical bounds and practical insights for improving the reliability of AI models in real-world scenarios with distribution shifts.
RANK_REASON The cluster contains a single academic paper published on arXiv detailing novel research findings. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- cs.LG
- natural-image corruption benchmark
- Prediction-powered inference
- Skeleton Action Recognition Based on Multi-Stream Spatial Attention Graph Convolutional SRU Network
- skeleton benchmark
- split conformal prediction
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
