Smooth Summaries of Persistence Diagrams and Texture Classification

Yu-Min Chung, Michael Hull, Austin LawsonWorkshops @ CVPR 2020 · DOI

Abstract

Persistent homology is a cornerstone of Topological Data Analysis (TDA), yet its primary output—persistence diagrams—often lacks the smoothness and Hilbert space structure required for modern machine learning. This paper introduces Gaussian Persistence Curves (GPCs), a novel framework that transforms these discrete diagrams into smooth, integrable functions. By replacing points with Gaussian kernels and integrating over filtration time, GPCs provide a summary that is both statistically robust and highly effective for complex tasks like texture classification. We provide a theoretical foundation through a Central Limit Theorem for GPCs and demonstrate state-of-the-art performance on benchmark texture datasets.

Introduction

Topological Data Analysis (TDA) captures the underlying "shape" of data through persistent homology. The resulting persistence diagram is a multiset of points representing the birth and death of topological features. However, these diagrams are difficult to integrate into standard machine learning pipelines like Support Vector Machines or Random Forests because they do not naturally reside in a Hilbert space. While existing summaries like persistence landscapes or images exist, they often lack smoothness or rigorous statistical properties. This paper addresses these gaps by proposing Gaussian Persistence Curves, which are smooth, easy to compute, and mathematically grounded.

Methodology

The GPC framework builds upon the existing Persistence Curve (PC) concept by introducing Gaussian smoothing and integration:

  1. Gaussian Smoothing: Each point in a persistence diagram is replaced with a 2D Gaussian kernel (a bell curve).
  2. Integration: This surface of Gaussians is integrated over a "fundamental box" associated with a filtration time $t$. This captures cumulative topological information in a way that respects the "Elder Rule" of persistent homology.
  3. Functional Output: The result is a smooth, 1D function of $t$. This function can be discretized into a vector for machine learning or used directly in functional data analysis.

Statistical Foundation

A major contribution of this work is the proof of a Central Limit Theorem (CLT) for GPCs. The authors demonstrate that as the sample size $n$ increases, the average GPC converges to a Gaussian process. This provides a solid theoretical footing for statistical inference, such as hypothesis testing and the construction of confidence intervals for topological features.

Convergence of GPC averages toward a Gaussian process as sample size increases.

Experimental Evaluation

The authors evaluated GPCs on both synthetic and real-world datasets:

  • Synthetic Shapes: Points sampled with heavy noise from a circle, a figure-8, and a "Swiss Roll" to test topological discrimination.
  • UIUCTex: A texture database containing 1,000 grayscale images across 25 classes.
  • KTH-TIPS2b: A challenging texture dataset with 810 images in 10 classes, featuring significant variations in illumination, pose, and scale.

Key Insights and Discussion

Smoothness and Robustness

By utilizing Gaussian kernels, the resulting curves are $C^1$ continuous and Lipschitz continuous. This inherent smoothness makes GPCs more robust to small perturbations in input data compared to "jagged" summaries like persistence landscapes. This stability is crucial when dealing with noisy real-world data.

Gaussian Persistence Curves for different topological spaces, demonstrating smoothness and discriminative power.

Bridging TDA and Statistics

Most TDA tools are descriptive. By proving weak convergence to a Gaussian process, the authors move GPCs into the realm of inferential statistics. This allows researchers to move beyond simple visualization and toward rigorous statistical comparisons of topological structures.

Key Results

Gaussian Persistence Curves proved highly competitive with state-of-the-art TDA methods:

  • Texture Accuracy: On the UIUCTex dataset, GPCs combined with Persistence Statistics (PS) and a Random Forest classifier achieved 94.1% accuracy.
  • Superior Robustness: On the KTH-TIPS2b dataset, the model achieved 94.7% accuracy, significantly outperforming Persistence Images (86.3%) and the Euler Characteristic Curve (89.4%).
  • Complementary Features: The best results were obtained by combining GPCs with traditional statistical measures of feature lifespans (mean, variance, entropy), suggesting that GPCs capture unique topological information.

Conclusion and Future Work

While GPCs offer a powerful and smooth summary of topological data, challenges remain regarding hyperparameter sensitivity, particularly the choice of the Gaussian bandwidth $\sigma$. Future research could focus on automated bandwidth selection and extending the framework to 3D data (e.g., medical imaging) or time-series analysis. Nonetheless, GPCs represent a significant step forward in making TDA both statistically rigorous and practically applicable for machine learning.

Summary updated

BibTeX
@inproceedings{chung2020smooth,
  author = {Yu-Min Chung and Michael Hull and Austin Lawson},
  title = {{Smooth Summaries of Persistence Diagrams and Texture Classification}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  doi = {10.1109/cvprw50498.2020.00428},
}
Figure