MAPLE: Self-supervised Learning-Enhanced Nonlinear Dimensionality Reduction for Visual Analysis

We present a new nonlinear dimensionality reduction method, MAPLE, that enhances UMAP by improving manifold modeling. MAPLE employs a self-supervised learning approach to more efficiently encode low-dimensional manifold geometry. Central to this approach are maximum manifold capacity representations (MMCRs), which help untangle complex manifolds by compressing variances among locally similar data points while amplifying variance among dissimilar data points. This design is particularly effective for high-dimensional data with substantial intra-cluster variance and curved manifold structures, such as biological or image data. Our qualitative and quantitative evaluations demonstrate that MAPLE can produce clearer visual cluster separations and finer subcluster resolution than UMAP while maintaining comparable computational cost.

💡 Research Summary

The paper introduces MAPLE, a novel nonlinear dimensionality‑reduction (DR) technique that builds on UMAP but addresses its most critical weakness: the construction of the weighted k‑nearest‑neighbor (k‑NN) graph from raw high‑dimensional data. In high‑dimensional spaces, Euclidean (or cosine) distances become concentrated, making the initial graph noisy and often misrepresenting the true geodesic relationships on curved manifolds. MAPLE replaces this static graph with a learned one by integrating multi‑view self‑supervised learning (MVSSL) and a Maximum Manifold Capacity Representation (MMCR) objective.

First, the method creates multiple augmented “views” of each data point (e.g., random masking, cropping, rotation) and feeds them through an encoder‑projector neural network. The network is trained to produce consistent embeddings across views, thereby learning a distance metric that reflects semantic similarity rather than raw pixel or gene‑level differences.

Second, the MMCR loss, derived from manifold‑capacity theory, simultaneously compresses the variance of locally similar points (by minimizing the nuclear norm of local embedding matrices) and expands the variance between dissimilar points (by maximizing the nuclear norm of inter‑cluster centroids). This dual pressure reshapes the embedding space so that a subsequent k‑NN graph more faithfully captures the underlying manifold geometry.

After this self‑supervised phase, MAPLE constructs a weighted k‑NN graph in the learned embedding space, converts it to a fuzzy graph, and finally runs UMAP’s standard cross‑entropy layout optimization. Because the graph now encodes a learned, data‑driven distance, the layout stage requires fewer corrective forces, leading to clearer cluster separation, reduced spurious overlaps, and finer sub‑cluster resolution.

The authors evaluate MAPLE on a variety of benchmarks, including single‑cell RNA‑seq data, image classification sets (CIFAR‑10, MNIST), and synthetic manifolds with high intra‑cluster variance. Quantitative metrics—K‑NN preservation, trustworthiness, silhouette score, and adjusted Rand index—show consistent improvements over UMAP, t‑SNE, PaCMAP, and recent UMAP‑derived variants (e.g., DensMAP). Qualitative visualizations illustrate that MAPLE can untangle densely packed clusters and reveal subtle structures that are blurred or merged in competing methods.

Computationally, MAPLE adds an upfront neural‑network training step, but with GPU acceleration the total runtime remains comparable to vanilla UMAP for datasets up to several hundred thousand points. The paper acknowledges memory scalability as a limitation for million‑point data and suggests future work on lightweight encoders or stochastic graph sampling.

In summary, MAPLE demonstrates that learning the neighborhood graph via self‑supervised representation learning and manifold‑capacity regularization can substantially improve the fidelity of graph‑based DR. By bridging modern SSL techniques with classic manifold‑preserving objectives, MAPLE offers a powerful, general‑purpose tool for visual analytics of high‑dimensional, complex‑structured data.