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

OpenViewer: Openness-Aware Multi-View Learning

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Uncertainty-aware multi-view learning · 6 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 OpenViewer, an openness-aware multi-view learning framework that combines pseudo-unknown sample generation (Mixup-style), an ADMM-derived deep unfolding network with interpretable prior modules, and an open-set loss regime, evaluated on six multi-view datasets with OSCR.

Problem

Multi-view learning models assume closed-set categories and are black boxes, so they lack interpretability of multi-view integration and fail to recognize unknown categories, often assigning them high-confidence known labels.

Approach

Generate pseudo-unknown multi-view samples by Mixup across classes, build an interpretable deep unfolding network from ADMM iterations with redundancy, consistency dictionary, diversity noise, and complementarity fusion modules, and train with losses that boost known-class confidence and suppress unknown-class confidence.

Key results

• On ESP-Game at openness 0.1, OpenViewer reaches CCR 58.65% at FPR 10% versus 40.05% for TMC (best baseline at that point), with the DN-Module lifting OpenSDViewer's ESP-Game result from 42.56% to 58.65%.
• OpenViewer achieves the best CCR at all listed FPR levels on Animals, VGGFace2-50, ESP-Game and NUSWIDE20k in the reported tables, e.g. 25.89% vs 7.96% for OpenSDViewer at 0.5% FPR on ESP-Game.
• OpenViewer is trained for 100 epochs with L=1 unfolding layer on a single RTX 4080 GPU.

Measured effects reported in the paper

Surprising findings

• On NUSWIDEOBJ and AWA at low FPR, TMC or RCML outperform OpenViewer in some cases.
• Increasing lambda_2 too much makes the model collapse, with all samples receiving meaningless confidences across classes.

Limitations

• Baselines are backbone networks from other multi-view tasks rather than open-set multi-view methods, since few exist.
• Theoretical boundedness and generalization proofs rely on assumed convergence and Lipschitz bounds, and the paper's proofs are loosely argued.
• Heterogeneous or incomplete multi-view data are not addressed.

Experimental setup

Datasets: Animals AWA NUSWIDEOBJ VGGFace2-50 ESP-Game NUSWIDE20k

Benchmarks: Animals AWA NUSWIDEOBJ VGGFace2-50 ESP-Game NUSWIDE20k

Metrics: OSCR CCR at FPR 0.5%/1%/5%/10%/50%

Baselines: MvNNcor TMC MMDynamics IMvGCN LGCN-FF ORLNet RCML OpenSViewer OpenSDViewer

Builds on: Mixup ADMM deep unfolding networks OSCR

Closest papers by meaning

• OpenView: Empowering MLLMs with Out-of-view VQA (NeurIPS 2026 · similarity 0.85)
• Rethinking Multi-view Representation Learning via Distilled Disentangling (CVPR 2024 · similarity 0.85)
• Deep Incomplete Multi-view Learning via Cyclic Permutation of VAEs (ICLR 2025 · similarity 0.84)
• EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.84)
• Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.84)
• Swiss Army Knife: Synergizing Biases in Knowledge from Vision Foundation Models for Multi-Task Learning (ICLR 2025 · similarity 0.84)