Introduction
Multi-view learning (MVL) leverages correlations across diverse data sources—such as combining image features with text descriptions—to enhance predictive performance. However, traditional MVL models often operate under a "closed-set" assumption, failing when encountering "unknown" classes in real-world deployments. This phenomenon, termed "agnostophobia," leads models to misclassify novel samples into known categories with high confidence. Furthermore, the "black-box" nature of many deep MVL models limits their utility in safety-critical applications where interpretability is paramount.
Proposed Methodology
OpenViewer addresses these challenges through a three-pronged openness-aware strategy designed to bridge the gap between controlled research environments and the "open" real world.
1. Pseudo-Unknown Sample Generation
To mitigate agnostophobia, the framework simulates unknown samples during training. Inspired by the "Mixup" technique, the authors generate pseudo-unknown samples by interpolating features from different known categories. This allows the model to "practice" encountering data that does not conform to its predefined buckets, improving its ability to reject truly unknown samples during testing.
2. Expression-Enhanced Deep Unfolding Network (EDUN)
The core of OpenViewer's interpretability lies in its deep unfolding architecture. Rather than using standard black-box layers, the network is derived from the iterations of the Alternating Direction Method of Multipliers (ADMM) optimization algorithm. This process "unfolds" the mathematical optimization into four distinct, functional modules:
- RF-Module: Eliminates redundant features within individual views.
- CD-Module: Ensures cross-view consistency via dictionary learning.
- DN-Module: Filters out noise and outliers.
- CW-Fusion: Dynamically weights views based on their class-separability (complementarity).
3. Perception-Augmented Training Regime
The training loss is designed to optimize the feature space for open-set recognition. It employs a distance margin and center loss to compress known class features into tight, distinct clusters while explicitly penalizing high-confidence scores for pseudo-unknown samples. This effectively manages entropy, ensuring "sharp" predictions for knowns and high uncertainty for unknowns.
Experimental Evaluation
The framework was evaluated across six diverse datasets, including Animals, AWA, NUSWIDEOBJ, and multi-modal datasets like ESP-Game. Using the Open-Set Classification Rate (OSCR) as the primary metric—which measures the trade-off between identifying knowns and rejecting unknowns—OpenViewer consistently outperformed state-of-the-art competitors such as TMC, RCML, and ORLNet.
Ablation studies confirmed the necessity of the interpretable modules; removing the noise-processing (DN) or dictionary-learning (CD) components resulted in significant performance degradation. Visualizations using t-SNE further demonstrated that OpenViewer creates a clear "no-man's land" for unknown samples, facilitating reliable rejection compared to traditional methods.
Conclusion
OpenViewer represents a significant step toward making multi-view learning robust for real-world applications. By grounding network architecture in mathematical optimization and explicitly accounting for unknown classes, it provides both transparency and superior open-set generalization. Future research will explore handling heterogeneous data structures and incomplete data where certain views may be missing.
Summary updated