Abstract
Manifold alignment (MA) is a critical technique for finding commonalities between disparate data domains, such as matching gene expressions to protein levels. However, traditional MA methods are often transductive, meaning they cannot easily incorporate new, unseen data without re-running the entire algorithm. This paper proposes a "guided" framework using geometry-regularized twin autoencoders to perform manifold alignment that supports out-of-sample extension. By regularizing neural networks with pre-computed manifold embeddings, the model learns to map new data into a shared latent space and enables direct translation between modalities.
Introduction
Most manifold learning techniques, including t-SNE and Isomap, are non-parametric. This poses a significant challenge in clinical or real-time settings: when a new sample arrives, the entire manifold must be recalculated, which is computationally expensive and risks data leakage. Furthermore, existing methods often struggle with cross-domain mapping—the ability to accurately predict a corresponding point in Domain Y given a point in Domain X.
Researchers at Brigham Young University address these limitations by wrapping traditional manifold alignment within a neural network framework. Their approach, "Guided Manifold Alignment," creates a parametric mapping that generalizes to new samples while preserving the intrinsic geometry of the original data.
Methodology
The core of the proposed framework is the Geometry-Regularized Twin Autoencoder. The process follows four primary steps:
- The Guide: A traditional MA method (such as MASH or JLMA) is used to generate a "gold standard" shared embedding for a training subset.
- Twin Autoencoders: Two independent autoencoders ($AE_X$ and $AE_Y$) are established for each respective domain.
- Multi-task Loss: The encoders are trained using a three-part loss function: Reconstruction Loss (preserving domain-specific features), Alignment Loss (mapping inputs to the "guide" embeddings), and Anchor Loss (ensuring known correspondences map to the same latent coordinates).
- Cross-Modal Translation: By performing "decoder swapping," the model can translate samples between domains. For instance, a sample from Domain X is passed through Encoder X and then through Decoder Y to predict its representation in Domain Y.
Experimental Results
The authors evaluated the framework using 42 datasets from the UCI Machine Learning Repository and real-world clinical data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).
Out-of-Sample Consistency
Using the Mantel test—a spatial correlation statistic—the authors demonstrated that the twin AE model successfully preserved the geometric relationships of the original manifold alignment for unseen data. The choice of the "guide" method and the alignment weighting parameter ($\lambda$) were found to be critical for performance.
As shown in the figure above, the model maintains high correlation with the ground truth manifold alignment across various methods, particularly when $\lambda > 0$. Using JLMA as a guide, the architecture achieved a Mantel correlation of $0.80 \pm 0.32$ on unseen test data.
Clinical Case Study: Alzheimer’s Disease
In the ADNI case study, the model was tasked with translating cognitive test scores (ADAS-Cog 13) to functional scores (FAQ). This is particularly useful in clinical settings where only one type of assessment may be available. The model successfully translated these scores with high accuracy, yielding an RMSE of approximately 0.9–1.0 on a 5-point scale.
Crucially, the "extension" accuracy for new patients was nearly identical to the original model's accuracy, proving that the learned mapping is robust and generalizes well to out-of-sample data.
Discussion and Conclusion
The Guided Manifold Alignment framework effectively bridges the gap between non-parametric manifold learning and parametric neural networks. By leveraging autoencoders for regularization, the model avoids overfitting and captures the intrinsic geometry of the data.
While the current work focuses on two-domain alignments and is dependent on the quality of the initial "guide" method, it provides a scalable path forward for multi-modal data integration. Future research will explore extending this architecture to $N > 2$ domains and integrating more sophisticated geometric guides to further enhance mapping fidelity.
Summary updated