Learning Riemannian Metrics for Classification of Dynamical Models

Fabio Cuzzolin, Stefano Soatto2005

TL;DR

This paper introduces a novel method for learning optimal distance metrics for classifying dynamical models. By leveraging differential geometry and a volume element minimization objective, they demonstrate significant improvements in classification accuracy for tasks like human action and identity recognition compared to existing methods.

What Problem Does This Paper Address?

Classifying complex systems, such as human movements, often involves representing them as dynamical models. A crucial step in this classification is defining a suitable "distance" between these models. Standard approaches rely on predefined distance functions, but no single function works best for all classification problems. This paper tackles the challenge of learning an optimal distance metric tailored to a specific dataset and classification task, aiming to improve the performance of simple nearest-neighbor classifiers.

Methods

The core of the paper's contribution lies in its use of differential geometry to learn a Riemannian metric. They propose a method that:

  • Leverages the Fisher Information Tensor: This tensor, derived from the probability distribution of the dynamical models, provides a natural Riemannian metric on the space of these models.
  • Introduces Pullback Metrics: They define a family of new metrics by "pulling back" the Fisher metric through a parameterized diffeomorphism (a smooth, invertible transformation). This allows them to warp the original metric space to better suit the data distribution.
  • Minimizes the Volume Element: The optimization objective is to find the parameters of the diffeomorphism that minimize the volume element of the induced pullback metric. The intuition is that by minimizing the volume element, the geodesics (shortest paths) of the learned metric will be forced to pass through the regions where the data is most densely clustered, thus improving separation between classes.
  • Applies to Linear Dynamical Models: The method is demonstrated on classes of linear dynamical systems, specifically AR(2) and ARMA models, which are relevant for representing time-series data like motion sequences.

Datasets

  • Mobo Database: This dataset, containing 600 image sequences of 25 people walking under various conditions (speed, slope, carrying a ball) from different viewpoints, was used for experiments. The sequences were simplified into scalar signals representing the silhouette width, which were then used to generate linear dynamical models.

Key Insights and Ideas

  • Metric Learning is Crucial for Classification: The paper strongly suggests that the choice of distance metric significantly impacts classification performance, and learning a metric tailored to the data is more effective than using generic ones.
  • Volume Minimization as a Proxy for Classification: The insight that minimizing the volume element of a learned metric can lead to better classification performance is a key contribution. This provides a computationally tractable objective function that correlates with improved class separation.
  • Diffeomorphisms as Powerful Tools: Using parameterized diffeomorphisms to warp the Fisher metric space offers a flexible way to explore a rich family of potential metrics, allowing for adaptation to the specific data distribution.
  • Global Embedding vs. Local Methods: The learned metric provides a global embedding of the data space, which is an advantage over some spectral methods that only provide local embeddings.

Key Results

  • Improved Classification Accuracy: The learned pullback metrics consistently outperformed classical distances (like Frobenius norm, gap metric, v-gap metric, subspace angles) and even the original Fisher geodesic distance in both identity and action recognition tasks.
  • Identity Recognition: While a challenging task with limited information, the learned metric showed a clear advantage, with performance improving as the training set size increased.
  • Action Recognition: The method demonstrated superior performance across various viewpoints and training set sizes, with the pullback metric consistently outperforming the best competitor. The paper visually shows how the learned diffeomorphism effectively separates clusters of different actions.

Limitations and Future Work

  • Focus on Linear Models: The current work is primarily focused on linear dynamical models. The authors acknowledge that extending this approach to more complex, non-linear models is a natural and important next step.
  • Computational Complexity: While the optimization problem is well-defined, finding the optimal diffeomorphism parameters can be computationally intensive, especially for high-dimensional systems.
  • Choice of Diffeomorphism Family: The performance is influenced by the choice of the family of diffeomorphisms. Exploring different families or more general deformation strategies could lead to further improvements.
  • Realistic Scenarios: The authors suggest that testing this approach in more realistic scenarios with richer features and more complex dynamics would be crucial for practical applications.

Summary updated

Abstract

Consider the problem of classifying motions, encoded as dynamical models of a certain class. Standard nearest neighbor classification then reduces to find a suitable distance function in the space of the models. In this paper we present a supervised differential-geometric method to learn a Riemannian metric for a given class of dynamical models in order to improve classification performances. Given a training set of models the optimal metric is selected among a family of pullback metrics induced by the Fisher information tensor through a parameterized diffeomorphism. Experimental results concerning action and identity recognition based on simple scalar features are shown, proving how learning a metric actually improves classification rates when compared with Fisher geodesic distance and other classical distance functions.
BibTeX
@misc{cuzzolin2005learning,
  author = {Fabio Cuzzolin and Stefano Soatto},
  title = {{Learning Riemannian Metrics for Classification of Dynamical Models}},
  year = {2005},
}
Figure