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

Multiple Instance Verification

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Loss design and classifier theory · 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.

Introduces multiple instance verification (MIV) and cross-attention pooling (CAP), a query-conditioned MI pooling method with distance-based and variance-excited attention, evaluated on QMNIST, signature, and FEVER verification tasks.

Problem

Verifying a query instance against a bag of target instances with heterogeneous, unknown relevance; naive combinations of MIL pooling and Siamese networks perform poorly and fail to identify key instances.

Approach

Cross-attention pooling makes the bag representation depend on the query (query-key-value cross-attention with shared projections producing both Siamese outputs), plus two new attention functions (distance-based attention and variance-excited multiplicative attention) and multi-head co-excitation to make similar instances distinguishable.

Key results

• On QMNIST, CAP-DBA-L2 reaches 0.742 AUROC and 0.680 accuracy versus 0.708 AUROC and 0.658 accuracy for the baseline.
• On signature verification, CAP-VEMA reaches 0.819 AUROC versus 0.756 for the best benchmark (PMA) and near-perfect key instance detection (i-AUROC 0.995).
• On QMNIST key instance detection, CAP-VEMA achieves avg. i-AUROC 0.832 versus 0.696 for the baseline and about 0.51 for GABMIL and PMA.

Surprising findings

• Benchmarks adapted from SOTA MIL methods (GABMIL, PMA, MSA) perform no better and sometimes significantly worse than a simple max-similarity baseline.
• Models can identify key instances incorrectly yet classify the bag correctly, and vice versa, so classification alone is insufficient to assess MIL models.

Limitations

• Key instance detection evaluated only on datasets with ground-truth key instance identifiers.
• Threshold selection for key instance prediction left out of scope because bag sizes vary.
• Classification accuracy on signature verification remains well below 100% despite near-perfect key instance detection.

Experimental setup

Datasets: QMNIST ICDAR 2011 SigComp signature verification data FEVER

Benchmarks: QMNIST ICDAR 2011 SigComp signature verification FEVER

Metrics: AUROC accuracy precision recall F1-score instance AUROC instance AP

Baselines: Baseline (max-similarity Siamese) GABMIL PMA MSA MI-Net Bi-LSTM

Builds on: Gated-Attention-Based MIL Pooling by Multi-head Attention (Set Transformer) Siamese networks transformer decoder cross-attention squeeze-and-excitation

Closest papers by meaning

• Loss-Based Attention for Deep Multiple Instance Learning (AAAI 2020 · similarity 0.84)
• Are Multiple Instance Learning Algorithms Learnable for Instances? (NeurIPS 2024 · similarity 0.84)
• EASE-DETR: Easing the Competition among Object Queries (CVPR 2024 · similarity 0.82)
• Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves (NeurIPS 2026 · similarity 0.81)
• Optimize Incompatible Parameters Through Compatibility-aware Knowledge Integration (AAAI 2025 · similarity 0.81)
• Mixture of In-Context Experts Enhance LLMs' Long Context Awareness (NeurIPS 2024 · similarity 0.81)