Research map / Label-efficient and robust learning / Recognition and long-tail
Global-Semantic Alignment Distillation for Partial Multi-view Classification
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 GLAD, an imputation-free self-distillation method for partial multi-view classification that uses ground-truth labels as semantic anchors to build teacher "ideal embeddings" and distills a margin-aware student that runs without labels.
Problem
Partial multi-view classification with missing views suffers from information redundancy and semantic misalignment across views; existing imputation-free methods rely on heuristic view-informativeness weighting, and imputation-based methods add noise and redundancy.
Approach
Label-guided teacher fuses label embeddings with partial multi-view token embeddings via masked cross-attention to produce ideal embeddings; a label-free student is trained to match them with an entropy-weighted (margin-aware) MSE distillation loss, plus classification loss.
Key results
- At missing rate 0, GLAD reaches 83.50% on Scene15 (+3.21% over the second-best method) and 83.38% on LandUse21 (+5.38%).
- At missing rate 0.5, GLAD reaches 59.62% on LandUse21 (+4.57%) and 66.69% on Scene15 (+2.92%) over the second-best method.
- Ablation at missing rate 0 shows the margin-aware weighting adds gains (e.g., Caltech101-20 95.18 to 96.44; LandUse21 83.33 to 84.52).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GLAD instead of the strongest reported baseline | PMvC classification · Scene15, Animal, BDGP, LandUse21, Caltech101-20; η = 0, 0.1, 0.5 | accuracy | Helped |
| Add global-semantic alignment distillation to the student network (ablation) | GLAD · LandUse21 and Caltech101-20, η = 0 | accuracy: 82.62 -> 83.33; 94.97 -> 95.18 | Helped |
| Add global-semantic alignment distillation to the student network (ablation) | GLAD · Scene15 and Animal, η = 0 | accuracy: 83.39 -> 83.84; 88.04 -> 88.53 | Helped |
| Add margin-aware weighting (ablation) | GLAD with global-semantic alignment distillation · Scene15, Animal, LandUse21, Caltech101-20, η = 0 | accuracy: 83.84 -> 83.95; 88.53 -> 89.12; 83.33 -> 84.52; 95.18 -> 96.44 | Helped |
Limitations
- UIMC runs out of memory on the Animal dataset, so comparisons there are incomplete.
- Hyperparameter lambda is tuned on a validation set and optimal values vary across datasets.
- Teacher relies on ground-truth labels during training, so it is not applicable without labels.
Experimental setup
Datasets: Scene15 Animal Caltech101-20 BDGP LandUse21
Benchmarks: Scene15 Animal Caltech101-20 BDGP LandUse21
Metrics: classification accuracy (ACC)
Baselines: Mean-Imputation CPM-Nets TMC Mmdynamics DD-IMvMLC-net DICNet UIMC RCML
Builds on: CPM-Nets knowledge distillation transformer attention
Closest papers by meaning
- CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (AAAI 2025 · similarity 0.88)
- Rethinking Multi-view Representation Learning via Distilled Disentangling (CVPR 2024 · similarity 0.86)
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.86)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.86)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.86)
- DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.