Reproduce “Are Independently Estimated View Uncertainties Comparable? Unified Routing for Trusted Multi-View Classification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yilin Zhang, Cai Xu, Haishun Chen, Ziyu Guan, Wei Zhao
TMUR separates evidential prediction from fusion arbitration in multi-view classification. Instead of comparing uncertainty values produced on incompatible view-specific scales, a cross-view router assigns sample-dependent weights to private and collaborative experts, improving both accuracy and calibration across heterogeneous benchmarks while remaining more robust to feature rescaling.
Key insights
- Positive rescaling of an evidence vector preserves its relative class support but lowers subjective-logic uncertainty, showing why independently generated uncertainties are not automatically comparable.
- Routing based on the joint multi-view context can access information unavailable to isolated branches and therefore better arbitrate which evidence sources should influence each sample.
- Router weights generally track expert reliability and remain comparatively stable under artificial view-feature rescaling, although they should not be interpreted as calibrated probabilities of correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.