📊 statistics

MEDAL: Manifold Embedding Distillation via Autoencoder Learning

The paper introduces MEDAL, a novel framework that distills static manifold embeddings into reusable autoencoder models to enable rigorous out-of-sample validation, quantitative distortion measurement, and hyperparameter tuning for nonlinear dimensionality reduction techniques.

Original authors: Irene Chang, Tarek M. Zikry, Genevera I. Allen

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Irene Chang, Tarek M. Zikry, Genevera I. Allen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, complex library of books (your high-dimensional data). To make sense of it, you want to create a simple, 2D map (an embedding) that shows how the books relate to one another. Tools like t-SNE and UMAP are popular cartographers that draw these maps. They are great at making beautiful pictures where similar books cluster together.

However, there is a big problem: these maps are static. Once the cartographer draws the map, they throw away the instructions on how to get there. If a new book arrives, you can't tell them where to put it on the map. Worse, you can't check if the map is actually accurate because you have no way to "reverse-engineer" the map to see if you can get back to the original book. You just have to look at the picture and hope it looks good. This is like trying to navigate a city using a hand-drawn sketch with no street names or compass.

MEDAL is a new framework that fixes this by turning that static sketch into a living, reusable GPS system.

Here is how it works, using simple analogies:

1. The "Teacher" and the "Student" (Knowledge Distillation)

Think of the original map (made by t-SNE or UMAP) as the Teacher. The Teacher knows exactly where every book in the library belongs on the 2D map, but it can't explain how it got there or handle new books.

MEDAL introduces a Student (a type of AI called an autoencoder). The Student's job is to watch the Teacher and learn to mimic its movements perfectly.

The Training: The Student looks at the original books and the Teacher's map. It tries to draw a line from the book to the map location.
The Constraint: The Student is strictly told: "You must land on the exact same spot the Teacher chose."
The Twist: Once the Student learns to place the book on the map, it must also try to rebuild the original book from that map location.

2. The "Reconstruction Test" (The Gold Standard)

This is the magic part. Because the Student has learned to rebuild the book from the map, we can now test it with new, unseen books (held-out data).

We take a new book, send it through the Student to get its spot on the map.
Then, we ask the Student to rebuild the book from that spot.
The Score: If the rebuilt book looks blurry or missing pages, we know the map distorted that book's information. If the book is rebuilt perfectly, the map is accurate.

This turns the map from a pretty picture into a testable scientific model. Instead of guessing if the map is good, we have a number (reconstruction error) that tells us exactly how much information was lost.

3. What MEDAL Helps You Do

The paper shows that this "GPS system" allows scientists to do four specific things that were previously impossible or just guesses:

Tuning the Dials (Hyperparameter Selection): When you use t-SNE or UMAP, you have to choose settings (like "perplexity"). Usually, people just pick the setting that makes the prettiest picture. MEDAL lets you pick the setting that preserves the most information for new data, like tuning a radio to the clearest signal rather than the loudest one.
Finding the "Blurry Spots" (Distortion Analysis): The system can point out exactly which books (or data points) are hard to fit on the map. In the paper's biological examples, this highlighted specific types of cells that are so complex or unique that squishing them into 2D makes them look "blurry" or distorted.
Spotting Imposters (Distribution Shift): Imagine you built a map of New York City. If you suddenly drop a book about Tokyo onto that map, the GPS system will scream, "This doesn't fit!" MEDAL can detect when new data (like cells from a different patient or a different lab) doesn't belong in the existing map, alerting scientists that something has changed.
Comparing Apples to Oranges: Before, comparing a t-SNE map to a UMAP map was like comparing a watercolor painting to a photograph—you couldn't really say which was "better" because they used different rules. MEDAL forces both methods to play by the same rules: "Who can rebuild the original data best?" This gives a fair, quantitative score to decide which method is superior for a specific dataset.

The Bottom Line

MEDAL doesn't replace the popular tools like t-SNE or UMAP. Instead, it wraps them in a quality control suit. It takes the static, untestable maps these tools create and turns them into dynamic, testable models. This ensures that when scientists make discoveries based on these maps—whether in cancer research, neuroscience, or astronomy—they are relying on a map that has been rigorously checked for accuracy, not just one that looks nice.

Technical Summary: MEDAL (Manifold Embedding Distillation via Autoencoder Learning)

Problem Statement

Low-dimensional embeddings are fundamental tools in modern scientific analysis for visualizing high-dimensional data and guiding downstream discoveries (e.g., in single-cell genomics and neuroscience). However, popular nonlinear dimension reduction (DR) methods like t-SNE, UMAP, and Isomap suffer from a critical validation gap. Unlike linear methods such as Principal Component Analysis (PCA), these nonlinear techniques typically provide embeddings only for the training data, lacking a general out-of-sample map or an inverse reconstruction to the original feature space. Consequently, they cannot be evaluated using held-out validation—the gold standard in supervised learning. This forces researchers to rely on visual appeal and qualitative inspection for method selection and hyperparameter tuning, leading to fragile workflows where small changes in initialization or parameters can drastically alter scientific conclusions. Existing diagnostic tools (e.g., EMBEDR, scDEED, neMDBD) offer pointwise reliability scores or local continuity metrics but remain task-specific, lack a unified held-out validation protocol, and do not provide a direct measure of information loss in the original feature space.

Methodology

The authors propose MEDAL (Manifold Embedding Distillation via Autoencoder Learning), a model-agnostic framework that converts a static, fitted "teacher" embedding into a reusable, parametric "student" model. MEDAL does not replace existing DR methods; rather, it acts as a validation wrapper that distills any pre-computed embedding (from t-SNE, UMAP, PCA, etc.) into a constrained autoencoder.

Core Mechanism

Teacher-Student Setup: A teacher embedding $Z_{tr}$ is first generated from a training set $X_{tr}$ using any chosen DR method.
Constrained Autoencoder Training: MEDAL trains a student autoencoder consisting of an encoder $e_\theta$ and a decoder $d_\phi$ . The training objective is a Lagrangian relaxation of two coupled losses:
- Distillation Loss ( $L_{dist}$ ): Forces the encoder's bottleneck output to match the teacher coordinates $Z_{tr}$ exactly (or within a small tolerance $\epsilon$ ).
- Reconstruction Loss ( $L_{rec}$ ): Minimizes the error between the original input $X_{tr}$ and the decoder's output $d_\phi(e_\theta(X_{tr}))$ .
The optimization problem is formulated as:
$\min_{\theta, \phi} \lambda_d L_{dist}(X_{tr}, Z_{tr}; \theta) + L_{rec}(X_{tr}; \theta, \phi)$
where $\lambda_d$ is a penalty weight tuned to ensure the distillation loss reaches a near-zero threshold.
Architectural Constraints: To ensure the reconstruction error is interpretable solely as a property of the teacher embedding (and not fitting artifacts), the authors enforce specific architectural choices:
- Linear Bottleneck: The mapping into the bottleneck layer is linear to avoid imposing artificial range restrictions on the teacher coordinates.
- Nonlinear Hidden Layers: Hidden layers use ReLU or SELU activations to provide the expressiveness needed to match complex, nonlinear teacher manifolds.
- Linear Output: The final reconstruction layer is linear to accommodate the full input space.

Validation Protocol

Once the student is trained to faithfully reproduce the teacher embedding (near-zero distillation loss), the encoder-decoder pair enables four downstream tasks:

Hyperparameter Selection: Candidate hyperparameters for a DR method are selected by minimizing the held-out reconstruction error on a validation set, rather than by visual inspection.
Pointwise Distortion Analysis: Reconstruction error is calculated for each observation, identifying specific regions of the manifold where the low-dimensional compression fails to preserve high-dimensional structure.
Distribution Shift Detection: New samples can be mapped to the fixed reference manifold. Samples with systematically high reconstruction errors are flagged as out-of-distribution or poorly represented by the reference geometry.
Cross-Method Comparison: Different DR methods can be compared on a shared scale by evaluating their held-out reconstruction errors under the same student architecture and validation split.

Key Results

The authors evaluate MEDAL across four case studies: the MNIST benchmark, Hydra single-cell RNA-seq, mouse neocortex single-cell RNA-seq, and macaque retina single-cell RNA-seq.

Hyperparameter Selection: On MNIST and biological datasets, MEDAL-selected hyperparameters (e.g., UMAP $n$ -neighbors, t-SNE perplexity) consistently balanced local separation with global coherence. In contrast, visually "sharp" embeddings often overfit training structure and failed to generalize to held-out data, while overly smooth embeddings compressed distinct classes. MEDAL successfully identified intermediate settings that preserved structure for unseen data.
Distortion Analysis: MEDAL identified specific cell populations with high reconstruction errors in biological datasets. For example, in the Hydra dataset, high-error regions consistently localized to interstitial nematocytes and male germline cells. In the mouse neocortex, non-neuronal and endothelial populations showed higher distortion than major neuronal classes. These findings were consistent across different teacher hyperparameters, suggesting MEDAL detects stable biological signatures of distortion rather than visualization artifacts.
Distribution Shift: In the macaque retina study, MEDAL successfully detected subject-level distribution shifts. When cells from a new subject (M2 or M3) were projected onto a manifold trained on a reference subject (M1), they exhibited spatially coherent increases in reconstruction error, whereas in-distribution cells remained low-error. This allowed for the localization of mismatch to specific cell types (e.g., microglia, endothelial cells).
Comparison with Existing Methods: MEDAL was compared against EMBEDR, scDEED, and neMDBD. MEDAL selected more balanced hyperparameters (avoiding extreme local or global regimes) and provided more interpretable, spatially coherent distortion maps. Furthermore, MEDAL demonstrated superior computational efficiency, completing hyperparameter sweeps significantly faster than permutation-based or null-distribution methods.
Reconstruction Validity: Experiments using PCA as a teacher confirmed that the distillation constraint does not introduce artificial reconstruction artifacts. Linear MEDAL exactly matched PCA reconstruction curves, and nonlinear MEDAL remained competitive with unconstrained autoencoders.

Significance and Claims

The paper claims that MEDAL addresses the "crisis in rigor and reproducibility" in scientific workflows by bringing the principles of model validation and selection to unsupervised dimension reduction.

Quantitative Validation: MEDAL converts static visualizations into testable models, enabling held-out validation for nonlinear embeddings. This allows researchers to select hyperparameters and methods based on quantitative information preservation rather than visual appeal.
Generalizability: As a wrapper, MEDAL is agnostic to the underlying DR method, providing a common evaluation scale for comparing disparate techniques (e.g., t-SNE vs. UMAP vs. PCA).
Interpretability: By linking reconstruction error directly to information loss in the original feature space, MEDAL offers a principled way to audit embeddings and identify biologically coherent regions that are difficult to represent in low dimensions.
Workflow Integration: The authors emphasize that MEDAL is designed to fit into existing scientific workflows, augmenting familiar tools like t-SNE and UMAP with rigorous validation steps before downstream analysis (clustering, trajectory inference, etc.).

The authors modestly note that MEDAL validates a chosen teacher embedding and inherits its assumptions; it does not guarantee that an embedding is scientifically "correct" if the teacher itself removes meaningful variation. Instead, it serves as a quantitative criterion within a broader scientific process, treating latent representations as hypotheses about structure that can be tested.