Imagine trying to wrap a flat sheet of paper around a donut or a sphere without creasing, tearing, or overlapping it. Mathematically, it is impossible. Yet, standard deep generative models—such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Normalizing Flows—attempt the high-dimensional equivalent of this feat every day. They take a simple, solid Euclidean latent space (like a Gaussian blob) and try to stretch it via a continuous neural network to fit real-world data.
When the underlying data manifold possesses a complex topology—such as holes, loops, or disconnected clusters—this single-mapping approach breaks down. As [Yinzhu Jin] et al. demonstrate in Implications of data topology for deep generative models, forcing a topologically simple latent bottleneck to map onto a non-trivial target space introduces severe geometric distortions, forcing the model to either tear the representation or fill topological holes with low-density artifacts.
To bypass these constraints, a growing body of research looks to classical differential geometry for a solution: the atlas. Instead of forcing data into a single global coordinate system, these methods decompose complex manifolds into a collection of overlapping, localized coordinate charts that are collectively glued together.
The Core Mechanism: Learning an Atlas
In differential geometry, an atlas describes a manifold through a collection of local charts, where each chart behaves like a flat Euclidean space. Where charts overlap, smooth transition functions ensure that moving from one coordinate patch to another preserves the manifold's global structure.
Applying this to machine learning means moving away from a "one-size-fits-all" global latent space. Early foundational work by [Chris Russell] et al. in Learning a Manifold as an Atlas showed that modeling manifolds as an ensemble of overlapping affine subspaces preserves non-trivial topologies like loops and spheres. By avoiding global flattening, this approach prevents the topological tearing and collapsing common in traditional manifold learning.
Global Mapping (Standard VAE/GAN) Atlas-Based Mapping (Multi-Chart)
[ Flat Latent Space ] [Chart 1] [Chart 2] [Chart 3]
│ \ │ /
(Forced Stretching) (Smooth Overlapping)
▼ ▼
[Torn or Distorted Manifold] [Continuous Complex Manifold]
Implementing Multi-Chart Generative Frameworks
To translate this geometric intuition into deep learning architectures, researchers have adapted various generative paradigms to support localized coordinate systems.
Localized GANs and VAE Mixtures
Instead of relying on a single global generator, models can distribute the workload across an ensemble of local experts. In Global Versus Localized Generative Adversarial Nets, [Guo-Jun Qi] et al. introduce an atlas of local generators. Each generator defines a local coordinate system around a specific data point, granting the model direct access to the manifold's local geometry while mitigating mode collapse.
Similarly, for inverse problems like image deblurring, [Giovanni S. Alberti] et al. propose a mixture of VAEs in Manifold Learning by Mixture Models of VAEs for Inverse Problems. Here, each individual VAE acts as a single chart within a broader atlas, allowing the network to capture disconnected components and non-Euclidean structures without introducing artificial gaps.
Multi-Chart Normalizing Flows
Normalizing flows are strictly continuous and bijective, making them exceptionally vulnerable to topological mismatches. To overcome this, [Janis Postels] et al. introduce Mixtures of Normalizing Flows in Go with the Flows: Mixtures of Normalizing Flows for Point Cloud Generation and Reconstruction. By partitioning complex geometries into local manifolds, the model represents sharp edges and gaps without requiring extremely deep architectures or extreme spatial stretching.
Expanding on this, [Hanlin Yu] et al. develop a framework in Learning geometry and topology via multi-chart flows that explicitly "glues" multiple local coordinate patches together. This multi-chart flow enables the calculation of shortest paths (geodesics) across surfaces that wrap around themselves, such as tori or spheres. The authors of ChartPointFlow for Topology-Aware 3D Point Cloud Generation similarly utilize unsupervised latent labels to segment point clouds into distinct charts, bypassing bijective constraints to cleanly generate holes and boundaries.
Gluing the Charts: Partitions of Unity and Local Trivializations
A collection of isolated charts is merely a set of fragments; to form a true manifold, they must be mathematically coordinated.
One prominent method for achieving this is detailed by [Jakob Stolberg-Larsen] et al. in Atlas Generative Models and Geodesic Interpolation. They implement a hybrid latent space that pairs discrete local charts with a partition of unity function. This mathematical tool smoothly blends the outputs of overlapping charts, allowing generative models to partition data into topological clusters while maintaining a continuous global representation.
[Chart A] ───► Local Coordinates ──┐
├──► [Partition of Unity Weighting] ───► Smooth Global Manifold
[Chart B] ───► Local Coordinates ──┘
When data exhibits complex global "twists" or many-to-many relationships, the manifold can be framed as a fiber bundle—a space that locally looks like a simple product space but globally possesses non-trivial connectivity. To model these structures, [Nico Courts] et al. introduce Bundle Networks in Bundle Networks: Fiber Bundles, Local Trivializations, and a Generative Approach to Exploring Many-to-one Maps. By partitioning the output space into local neighborhoods, the model performs local trivializations. This strategy breaks down global twists into simpler, localized mappings that collectively capture complex geometric loops.
Following a similar geometric philosophy, [Loc Truong] develops the Bundle Morphism Network in Fiber Bundle Morphisms as a Framework for Modeling Many-to-Many Maps, using conditioning vectors based on data clusters to define local neighborhoods and learn the specific geometric transformations inherent to different regions of the manifold.
Applications in 3D Surface Generation
The practical utility of atlas-based representations is most visible in 3D shape modeling, where objects naturally possess varying numbers of holes (genera) and boundaries.
- The Papier-Mâché Approach: In A Papier-Mâché Approach to Learning 3D Surface Generation, [Matthew A Fisher] et al. introduce AtlasNet. This architecture models 3D shapes as a collection of independent parametric patches trained to deform 2D squares into 3D space. This ensemble of local charts allows the model to generate complex topologies without requiring an explicit prior on the object's global genus.
- Spherical and Flat Parameterizations: For high-genus surfaces, [David Gu] et al. demonstrate in Conformal Spherical Parametrization for High Genus Surfaces that multi-layered spherical parameterizations can process complex topologies without introducing artificial seams. From a different angle, [Siming Li] shows in Space Filling Curves for 3D Sensor Networks with Complex Topology that a complex surface with multiple handles can be unrolled into a branched covering of connected flat parallelograms, enabling continuous path construction across topologically distinct regions.
Theoretical Justification
Why does decomposing a manifold into local regions work so effectively? [Biraj Dahal] et al. provide a rigorous theoretical foundation in On Deep Generative Models for Approximation and Estimation of Distributions on Manifolds. They prove that deep neural networks can approximate distributions on complex, non-flat manifolds by decomposing the manifold into local regions and integrating the resulting maps. Crucially, they demonstrate that the error in learning these distributions depends entirely on the intrinsic dimension of the data manifold rather than the high ambient dimension of the raw data space, effectively bypassing the curse of dimensionality.
Furthermore, this localization does not mean giving up control over the latent space's global structure. As [Ning Miao] et al. show in On Incorporating Inductive Biases into VAEs, specific mapping operations like sector-splitting or orthogonal projections can be injected directly into the framework. This allows researchers to enforce desired connectivity and modality constraints on the learned representations, combining the training stability of standard Gaussian frameworks with the topological flexibility of a multi-chart atlas.
Key Foundational Work
Learning a Manifold as an Atlas (2013) — contributed 2 concept elements to this direction.
Conformal Spherical Parametrization for High Genus Surfaces (2007) — contributed 1 concept elements to this direction.