Most machine learning models assume that data lives in a flat, Euclidean space. Whether it is a Variational Autoencoder (VAE) compressing images into a latent "blob" or a Normalizing Flow stretching a Gaussian distribution, these models attempt to find a single, global coordinate system for the data.
However, real-world data often lives on manifolds with complex topologies—shapes like spheres, tori, or loops that cannot be flattened without tearing. Just as you cannot flatten an orange peel onto a table without ripping it, you cannot map a spherical data distribution to a flat plane without creating mathematical artifacts. To solve this, researchers are turning to the concept of an Atlas: representing a complex global shape as a collection of overlapping local patches.
The Orange Peel Problem
The fundamental issue is what Leonie Papon identifies as "topological obstructions." In Quantitative Rates and Fundamental Obstructions to Non-Euclidean Universal Approximation, Papon proves that neural networks can only globally approximate functions that are "topologically simple." If the data has "holes" or wraps around itself, a single continuous mapping will inevitably fail.
C. Grace Haaf and colleagues demonstrate this in Topological Obstructions to Autoencoding, showing that standard autoencoders create "singularities" or artificial reconstruction errors when forced to represent manifolds like circles or spheres. These models try to "tear" the manifold to make it fit into a flat latent space, leading to "dead zones" where the model cannot generate valid data.
The Atlas Solution: Local Charts and Transition Maps
Instead of a single global map, an Atlas approach uses a collection of local coordinate "charts." Each chart covers a small neighborhood of the data and is topologically equivalent to a flat Euclidean space. This allows us to use standard neural network architectures locally while preserving the global structure.
Chris Russell and Lourdes Agapito pioneered this in Learning a Manifold as an Atlas, modeling manifolds as overlapping affine subspaces. The magic happens at the overlaps: transition maps ensure that as you move from one patch to another, the representation remains continuous.
This "patchwork" approach has been successfully applied to 3D shape generation. In A Papier-Mâché Approach to Learning 3D Surface Generation, Thibault Groueix and collaborators introduced AtlasNet, which generates complex 3D meshes by deforming a set of 2D squares (charts) and "gluing" them together in 3D space. This allows the model to represent objects with arbitrary genus (like a coffee mug with a handle) without needing to know the topology in advance.
Generative Modeling with Multi-Chart Flows
Normalizing flows are particularly sensitive to topology because they rely on bijective (invertible) mappings. A single bijection cannot map a flat space to a sphere. To overcome this, Dimitris Kalatzis and Søren Hauberg developed Multi-chart flows.
By using an atlas of local normalizing flows, the model can define probability densities on manifolds like tori or spheres. Hanlin Yu further extended this in Learning geometry and topology via multi-chart flows, providing algorithms to compute shortest paths (geodesics) across these patches. This allows for smooth interpolation between data points even when the path must "wrap around" a hole in the data.
Beyond Simple Surfaces: Fiber Bundles
Sometimes the "many-to-one" nature of data requires even more sophisticated geometry. Henry Kvinge and Nico Courts introduced Bundle Networks (BundleNets), which use the concept of fiber bundles.
In this framework, the input space is treated as a "total space" that projects down to a "base space" (the labels). The set of all inputs that map to the same label is called a "fiber." By using local trivialization—essentially an atlas for these fibers—BundleNets can accurately sample the entire distribution of inputs that lead to a specific output, outperforming standard conditional GANs.
Why It Works: Intuition
The atlas approach works because it decouples local computation from global topology:
- Local Simplicity: At a small enough scale, every smooth manifold looks like Euclidean space. We can use our most powerful, "flat" tools (like linear layers or standard flows) within these local charts.
- Global Connectivity: By enforcing consistency in the overlapping regions (via transition maps or shared latent variables), the model "glues" these local truths into a global whole.
- Topological Flexibility: Because the patches are independent, the model can represent disconnected components, holes, and self-intersections that would "break" a global mapping. Mete Özay shows in Compositional Hierarchical Representation of Shape Manifolds that this even allows for the classification of "non-manifold" shapes with disconnected parts.
As Ryan A. Robinett and Lorenzo Orecchia argue in their recent work on Atlas-based Manifold Representations, this shift from "squashing" data to "patching" it allows for machine learning that is not only more accurate but also more interpretable, as it respects the intrinsic Riemannian geometry of the data.