When we try to understand a complex object, we rarely rely on a single source of information. A doctor diagnoses a patient by looking at both medical images and clinical history; a social scientist profiles a user by looking at their activity across multiple social networks. In machine learning, these different sources are called "views."
The core intuition behind Multi-view Consensus Regularization is that while any single view might be noisy, incomplete, or biased, the underlying truth should be consistent across all of them. By forcing different models to agree on a single "opinion," we can filter out the noise unique to individual sensors or perspectives and arrive at a more robust, "trusted" prediction.
The Consensus Principle as a Denoising Filter
In a standard single-view model, the network is free to overfit to any pattern in the data that reduces the loss, even if that pattern is just random noise or a dataset-specific quirk. However, when we train multiple models on different views—such as the text of a webpage and its hyperlink structure—and penalize them for disagreeing, we create a powerful filter.
As Jim Jing-Yan Wang and Ivor W. Tsang argue in Optimizing Multivariate Performance Measures from Multi-View Data, cross-view consistency acts as a regularizer. Because noise is typically independent across views, it is unlikely to lead to the same incorrect prediction in both. Therefore, the only way for two models to consistently agree is to focus on the shared, invariant features of the target object.
From Predictions to "Trusted" Opinions
Recent work has moved beyond simple label agreement toward a more sophisticated "evidential" framework. Instead of just outputting a class probability, models are treated as agents providing "opinions" with varying levels of certainty.
Wei Liu and Thierry Denoeux formalize this in Trusted Multi-View Deep Learning with Opinion Aggregation. They use evidence theory to represent model outputs as Dirichlet distributions. This allows the system to quantify not just what it thinks, but how much evidence it has. When views conflict, the model doesn't just pick one; it uses a "conflict-aware" fusion rule to identify which view is more reliable.
This is particularly vital for fairness and safety. For instance, Haishun Chen and Cai Xu show in Fairness-Aware Multi-view Evidential Learning that standard models often "unfairly" allocate evidence to majority classes. By using an adaptive prior and alignment metrics, they ensure that minority classes receive balanced evidence across all views, preventing overconfident misclassifications.
Architectural Diversity as a View
One of the most flexible aspects of this research direction is what constitutes a "view." While it often refers to different modalities (like audio and video), it can also refer to different ways of processing the same data.
In Multi-View Consistency for Relation Extraction, Amir Pouran Ben Veyseh and Thien Huu Nguyen force agreement between a sequence-based model (which sees word order) and a graph-based model (which sees syntactic structure). Similarly, Acong Zhang uses a contrastive scheme between a Graph Convolutional Network and a simple MLP in Building Shortcuts between Distant Nodes.
By requiring these fundamentally different architectures to reach a consensus, the system learns features that are "architecture-invariant," making the final model much more robust to the specific biases of any single neural design.
Handling Incomplete and Noisy Data
The ultimate test for consensus-based learning is when data is missing. In many real-world scenarios, like traffic forecasting or medical diagnosis, one or more views might be unavailable.
Haojian Huang and Qiang Li address this in Towards Robust Uncertainty-Aware Incomplete Multi-View Classification. Their approach uses consistency loss to "fill in the gaps." If the model learns that View A and View B usually agree, it can use the evidence from View A to refine its internal representation of View B, even when View B's input is noisy or missing. This "cross-view adaptation" ensures that the model's performance doesn't collapse when the environment is less than perfect.