[Paper Review] Uncertainty-Aware Multi-View Representation Learning

Yu Geng, Zongbo Han+2 more|arXiv (Cornell University)|Jan 15, 2022

Domain Adaptation and Few-Shot Learning27 references4 citations

TL;DR

This paper proposes Dynamic Uncertainty-Aware Networks (DUA-Nets), a novel unsupervised multi-view representation learning framework that dynamically weights views based on sample-specific uncertainty estimates derived from variational autoencoder-style reconstruction. By modeling view-wise uncertainty as a byproduct of reconstruction variance, DUA-Nets adaptively suppress noisy views and enhance robustness, achieving state-of-the-art performance on multiple benchmarks with superior noise resilience.

ABSTRACT

Learning from different data views by exploring the underlying complementary information among them can endow the representation with stronger expressive ability. However, high-dimensional features tend to contain noise, and furthermore, the quality of data usually varies for different samples (even for different views), i.e., one view may be informative for one sample but not the case for another. Therefore, it is quite challenging to integrate multi-view noisy data under unsupervised setting. Traditional multi-view methods either simply treat each view with equal importance or tune the weights of different views to fixed values, which are insufficient to capture the dynamic noise in multi-view data. In this work, we devise a novel unsupervised multi-view learning approach, termed as Dynamic Uncertainty-Aware Networks (DUA-Nets). Guided by the uncertainty of data estimated from the generation perspective, intrinsic information from multiple views is integrated to obtain noise-free representations. Under the help of uncertainty, DUA-Nets weigh each view of individual sample according to data quality so that the high-quality samples (or views) can be fully exploited while the effects from the noisy samples (or views) will be alleviated. Our model achieves superior performance in extensive experiments and shows the robustness to noisy data.

Motivation & Objective

• To address the challenge of integrating multi-view data with varying noise levels and quality across samples and views in unsupervised representation learning.
• To develop a method that adaptively assigns higher importance to high-quality views per sample, rather than using fixed or equal weights.
• To estimate data uncertainty in a self-supervised manner without requiring labels, using reconstruction variance from generative models.
• To unify uncertainty estimation and representation learning into a single end-to-end framework that improves both components jointly.

Proposed method

• DUA-Nets employ Reversal Networks (R-Nets) to reconstruct each view from a shared latent representation, enabling joint learning of representation and uncertainty.
• For each view, the R-Net outputs both the mean and variance of a Gaussian distribution, where the variance serves as an uncertainty estimate for that view.
• Uncertainty is used to dynamically weight views during representation learning, giving more influence to low-uncertainty (higher-quality) views.
• The model jointly optimizes reconstruction loss and uncertainty-based weighting, enabling mutual improvement between representation quality and uncertainty estimation.
• A collaborative learning mechanism ensures that uncertainty estimation guides representation learning and vice versa, enhancing robustness.

Experimental results

Research questions

• RQ1 Can uncertainty estimation from reconstruction variance effectively guide dynamic view weighting in multi-view representation learning?
• RQ2 How does uncertainty modeling improve robustness to noisy views in an unsupervised setting?
• RQ3 Can the model distinguish between clean and noisy samples based on uncertainty without supervision?
• RQ4 How does the integration of uncertainty estimation affect the quality and generalization of learned representations?

Key findings

• DUA-Nets outperform standard R-Nets without uncertainty modeling on clustering tasks, demonstrating that uncertainty improves representation quality.
• The model maintains stable performance under increasing noise intensity, while baseline R-Nets degrade significantly, showing superior robustness.
• Uncertainty estimates effectively capture data quality differences: higher noise ratios lead to increased uncertainty, even when noise is synthetically injected.
• The model's uncertainty estimation remains stable and informative even when 50% of samples are corrupted by noise, indicating strong generalization.
• Optimal performance is achieved with a latent dimension of 50, showing that both under- and over-parameterization degrade performance.