Flat Metric Minimization with Applications in Generative Modeling

Thomas Möllenhoff, Daniel Cremers

Introduction

Standard generative models, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), typically operate under the Manifold Hypothesis, assuming high-dimensional data resides on a low-dimensional manifold. However, these models often treat data as a "cloud of points"—a 0-dimensional probability distribution—ignoring the inherent local geometry and directional information. For instance, a digit's appearance changes predictably under rotation or thickening, and video frames evolve according to the flow of time. Traditional point-cloud approaches struggle to incorporate these tangent vectors, often resulting in "tangled" latent spaces where individual variables lack clear physical interpretations.

Methodology: From Distributions to $k$-Currents

To address these limitations, researchers from the Technical University of Munich have introduced FlatGAN, a framework rooted in Geometric Measure Theory (GMT).

Geometric Representation

Instead of viewing data as a collection of points, FlatGAN treats it as a $k$-current. A $k$-current can be conceptualized as a generalized oriented manifold. By setting $k > 0$, the model can "attach" a $k$-dimensional tangent plane to every data point, capturing local geometric variations.

Conceptual framework of FlatGAN mapping latent space to data space using k-currents and tangent vectors.

The Flat Metric and Optimization

The core of the model is the minimization of the flat metric, a fundamental GMT distance measure between two currents. This metric generalizes the Wasserstein distance (used in WGANs) to oriented geometric objects. In the FlatGAN formulation, the discriminator acts as a "geometric critic." Beyond evaluating image realism, it utilizes a "pullback" mechanism to inspect the Jacobian of the generator. It ensures that the directions in which the generator stretches the latent space align with the tangent vectors provided in the training data. The authors proved that this flat metric is Lipschitz continuous with respect to model parameters, facilitating stable training via gradient descent.

Experimental Evaluation

FlatGAN was evaluated across several datasets to test its ability to learn equivariant and disentangled representations:

  • MNIST: Tangent vectors were defined for rotation and line thickness.
  • smallNORB: Tangents represented changes in lighting, elevation, and azimuth.
  • Tinyvideos (Beach): A dataset of video frames where the "tangent" was the difference between successive frames.

Key Insights and Results

FlatGAN achieves disentanglement by design. By explicitly providing directional information (e.g., "this is the direction of rotation"), the model is forced to align specific latent dimensions with those physical transformations. This semi-supervised approach "anchors" the latent space, making it more predictable than unsupervised methods like InfoGAN.

Latent Walks and Video Synthesis

On MNIST and smallNORB, moving along a single latent axis resulted in smooth, interpretable transformations, such as a digit rotating perfectly. On the Beach dataset, the model successfully discovered the "arrow of time." By using frame differences as tangent vectors, FlatGAN learned a latent variable representing temporal progression, allowing it to generate coherent motion sequences from shuffled static frames.

Latent walks on the Beach dataset showing synthesized video frames with a learned temporal variable.

Limitations and Future Directions

While FlatGAN introduces a powerful geometric formalism, it faces computational challenges. Working with the full space of $k$-vectors is expensive, requiring heuristic approximations for high-dimensional data. Additionally, the model requires pre-defined tangent vectors, which may be difficult to obtain for abstract datasets. Future work aims to scale these geometric principles to more complex architectures and improve the raw visual quality of the generated outputs.

Summary updated

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