Research map / Label-efficient and robust learning / Recognition and long-tail

Conveyance: A Versatile Framework for Learning in Structured Class Spaces

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Learning from unlabeled and positive-unlabeled data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Conveyance, a cross-entropy-like loss that uses a user-specified boolean class-plausibility matrix Q to add a second margin on the plausible-set probability mass, applied to label noise, MIL, ordinal age regression and hierarchical CUB classification.

Problem

Standard classification losses such as cross-entropy are class-symmetric and cannot exploit structural relations between classes under structured label noise, ordinal targets, or taxonomies.

Approach

A log-sum-exp loss over two margins: target-vs-non-target odds (weight alpha) and plausible-set-vs-implausible mass (weight beta), with plausibility encoded as a binary matrix Q; reduces to CE at alpha=1, beta=0, keeps CE's cost, and is monotone and partially convex in the logits.

Key results

• On CIFAR-10 under asymmetric noise, Conveyance reaches 89.68% vs 80.53% for CE and 83.54% for Masking.
• On Camelyon-16, Conveyance achieves AUC 0.977 vs 0.959 for the next-best method CAMIL.
• On CUB zero-shot OOD genus accuracy, Conveyance reaches 85.1% vs 84.1% for CE+SupCon and 78.0% for CE, at CE-level complexity.

Surprising findings

• On CIFAR-100 with asymmetric noise, Conveyance (70.43%) slightly exceeds the clean-data CE baseline (70.14%) reported in the table.

Limitations

• Performance depends on the quality of Q; misspecified or incomplete masks may reinforce spurious plausibility.
• Effective alpha strength scales with the number of classes, requiring per-dataset recalibration.
• Constructing Q for large-vocabulary settings (e.g., ImageNet) requires structured domain knowledge that may be unavailable.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 CUB-200-2011 Camelyon-16 MUSK1 MUSK2 FOX TIGER ELEPHANT CACD2000 CLAP2016 UTKFace

Benchmarks: CIFAR-10 CIFAR-100 Camelyon-16 MUSK1 MUSK2 FOX TIGER ELEPHANT CACD2000 CLAP2016 UTKFace CUB-200-2011

Metrics: top-1 accuracy TFLOPS accuracy AUC F1 MAE epsilon-error genus accuracy OOD genus accuracy

Closest papers by meaning

• Learning Structured Representations by Embedding Class Hierarchy (ICLR 2023 · similarity 0.86)
• Symmetry Reveals Layerwise Dynamics: How Transformers Perform In-Context Classification (ICML 2026 · similarity 0.85)
• Automated Loss function Search for Class-imbalanced Node Classification (ICML 2024 · similarity 0.84)
• Learning Task-Agnostic Representations through Multi-Teacher Distillation (NeurIPS 2025 · similarity 0.84)
• Trading off Consistency and Dimensionality of Convex Surrogates for Multiclass Classification (NeurIPS 2024 · similarity 0.84)
• I-Con: A Unifying Framework for Representation Learning (ICLR 2025 · similarity 0.83)