Flat Metric Minimization with Applications in Generative Modeling

Thomas Möllenhoff, Daniel Cremers

Introduction

Generative models have traditionally treated data as a simple probability distribution—a collection of points scattered in a high-dimensional space. However, this perspective often leaves critical structural information on the table. By viewing data through the lens of geometric measure theory, a new approach called FlatGAN represents data as "$k$-currents." Essentially, these are distributions equipped with attached "arrows" or tangent vectors. This novel formulation allows the generative model to learn disentangled latent representations that are explicitly equivariant to physical transformations, such as rotation, lighting changes, or the flow of time.

The Missing Geometry in Generative Models

Most standard generative models, like standard GANs or VAEs, rely on the "manifold hypothesis," which suggests that high-dimensional data concentrates near low-dimensional surfaces. While these models learn to map latent spaces to these surfaces, they often ignore the local geometric structure—specifically, how the data changes locally.

Consider a video: it is not merely a random bag of frames, but a sequence where one frame smoothly transforms into the next. Standard GANs struggle to capture this "arrow of time" or specific continuous variations (like the stroke thickness in handwriting) without relying on complex, often unstable, unsupervised techniques. There is a fundamental gap between learning pure probability distributions and leveraging the rich geometric information (tangent vectors) that is often implicitly available in our datasets.

Enter FlatGAN: Geometric Measure Theory in ML

To bridge this gap, the authors introduce FlatGAN, which reformulates generative modeling using Geometric Measure Theory (GMT).

  1. From Distributions to Currents: Instead of standard probability distributions (0-currents), FlatGAN utilizes $k$-currents. A $k$-current can be thought of as a distribution where every point has an associated $k$-dimensional oriented tangent plane.
  2. The Flat Metric: To compare these currents, the model employs the Flat Norm (originally proposed by Whitney in 1957). Unlike the standard "mass" of a manifold—which is discontinuous and poorly suited for gradient descent—the flat metric is continuous and metrizes "weak convergence." This provides a mathematically sound distance metric for training neural networks.
Conceptual overview of FlatGAN mapping a latent space with vector fields to a data manifold to match target data currents.
  1. A Geometric Discriminator: FlatGAN derives a primal-dual objective similar to the Wasserstein GAN (WGAN). However, the FlatGAN "discriminator" acts as a differential $k$-form. It doesn't just evaluate the generator's output points; it inspects the Jacobian (the derivatives) of the generator. This ensures that the generated "tangent planes" stretch and rotate to match the real data's local geometry.

Key Insights: Orientation and Equivariance

By "orienting" the data with tangent vectors, FlatGAN provides the model with a roadmap of how the latent space should map to the data space. If the model is instructed that a specific latent dimension corresponds to a specific tangent vector (e.g., rotation), it is forced to make that latent variable control that exact transformation. This equivariance by design leads to highly disentangled representations where different latent variables control distinct, interpretable physical properties.

Interestingly, this framework also reveals that the standard WGAN is simply a special case of FlatGAN where $k=0$ (no tangent information is used), providing a principled mathematical bridge to higher-dimensional geometric learning.

Experimental Results: Discovering the Arrow of Time

The authors evaluated FlatGAN across several datasets, including synthetic 2D circles, MNIST (using rotation and thickness tangents), and smallNORB (using lighting, elevation, and azimuth). In these tests, the model successfully separated complex physical transformations into distinct latent dimensions simply by being provided with the difference between images as tangent vectors during training.

One of the most striking results came from the tinyvideos dataset, which consists of over 36 million frames of beach scenes.

FlatGAN successfully learns a latent dimension corresponding to the flow of time in the tinyvideos dataset.

While standard WGAN-GP failed to find a meaningful temporal walk in the latent space, FlatGAN successfully learned a latent dimension that explicitly corresponds to the flow of time, capturing dynamic elements like waves crashing and people walking across the beach.

Limitations and Future Directions

While FlatGAN offers a robust mathematical framework (including proofs that the flat metric is Lipschitz continuous for gradient-based optimization), it does have limitations. Unlike purely unsupervised methods, it requires prior knowledge of tangent vectors—such as knowing which frames are sequential or how a digit rotates. Furthermore, computing Jacobians and multiple backpropagation passes makes it more computationally expensive than standard GANs, and it relies on heuristic "affine" approximations to remain tractable in high dimensions.

Nevertheless, FlatGAN represents a compelling first step in leveraging Geometric Measure Theory for machine learning, opening the door for future applications in manifold denoising and advanced topological data analysis.

Summary updated

Abstract

We take the novel perspective to view data not as a probability distribution but rather as a current. Primarily studied in the field of geometric measure theory, k-currents are continuous linear fu...
BibTeX
@inproceedings{mollenhoff2019flat,
  author = {Thomas Möllenhoff and Daniel Cremers},
  title = {{Flat Metric Minimization with Applications in Generative Modeling}},
  booktitle = {International Conference on Machine Learning},
  year = {2019},
  url = {https://openreview.net/forum?id=ry-Ew2WdZH},
}
Figure