Random Filters for Enriching the Discriminatory power of Topological Representations

Grayson Jorgenson, Henry Kvinge, Tegan Emerson, Colin C. OlsonTAG-ML 2022 · OpenReview

Abstract

Topological representations of data are inherently coarse summaries which endows them with certain desirable properties like stability but also potentially inhibits their discriminatory power relative to other representations. We propose a framework for enriching topological representations of images using random filters. By applying a set of random filters to an image and computing the persistence of the resulting "interference" between the image and the filter, we can generate a multi-channel topological representation that is significantly more discriminatory than standard topological summaries. We demonstrate the effectiveness of this approach on the MNIST dataset, achieving state-of-the-art performance for TDA-based methods while being orders of magnitude faster than previous structured filter approaches.

Introduction: The Coarseness of Topology

In machine learning, Topological Data Analysis (TDA) is often valued for its ability to capture the "big picture" shape of data. By focusing on fundamental properties like connected components and holes, TDA provides features that are remarkably stable against noise and minor deformations. However, this stability is a double-edged sword: it stems from a fundamental coarseness.

Consider the handwritten digits "0", "6", and "9". To a standard topological algorithm, these are often indistinguishable because they each consist of a single loop. They are topologically equivalent. This lack of fine-grained detail limits the discriminatory power of TDA in tasks like digit recognition (MNIST), where subtle geometric differences—the position of a loop, the length of a stroke—are more important than raw connectivity.

Methodology: Creating Interference Topology

To overcome this limitation, we introduce a framework that moves beyond capturing the "direct topology" of an image. Instead, we capture its "interference topology" using random filters.

1. Lower Star Filtration (LSF)

The standard approach for images is Lower Star Filtration. It treats pixel intensities as "heights" and builds a sequence of simplicial complexes (shapes) as one sweeps from the darkest to the brightest pixels. This captures the birth and death of topological features based on intensity.

2. Random Filters and the "Imprint"

Rather than analyzing the raw image $I$, we generate a set of "random filters" $R$. These are created by taking a matrix of uniform noise and applying a Gaussian blur to ensure some spatial correlation. We then compute the element-wise product (Hadamard product) of the image and the filter: $I' = R \odot I$.

Visualizing the interference topology: Random filters applied to digits '1' and '5' create distinct multi-channel representations.

This process can be thought of as looking at the digit through a textured glass pane. The digit's specific geometry "interferes" with the random patterns of the filter, creating a unique topological signature that highlights features otherwise invisible to standard TDA.

3. Multi-channel Persistence Images

For each filtered image, we generate a Persistence Diagram, which records the birth and death of topological features (e.g., connected components in Dimension 0).

A typical Persistence Diagram (Dimension 0) showing birth and death coordinates of topological features.

These diagrams are then vectorized into Persistence Images (PIs). By using $N$ different random filters, we create a multi-channel PI (similar to RGB channels in standard images), which serves as the input for a simple Convolutional Neural Network (CNN).

Key Insights: Why Randomness Works

A core finding of this work is that diversity outperforms structure. Previous attempts to enrich TDA used "structured filters" like edge detectors or radial gradients. We argue that random filters provide a more diverse set of "perspectives" on the data.

This is analogous to the Johnson-Lindenstrauss lemma in dimensionality reduction: just as random projections can preserve distances in high-dimensional space, random topological filters can preserve and highlight discriminatory features that a single "perfect" view might miss. Furthermore, random filters are computationally efficient; they require only a single noise generation and simple multiplication, avoiding the iterative processing required by many structured filters.

Results and Performance

We evaluated our approach on the MNIST dataset (60,000 training, 10,000 test samples). Using a consistent, simple CNN architecture, we compared random filters against raw images and structured filters.

  • Accuracy:
    • Raw MNIST Images: 91.16%
    • 25 Structured Filters: 95.33%
    • 25 Random Filters: 97.17%
    • 80 Random Filters: 98.2%
  • Computational Efficiency: Generating 25 random filters for 1,000 samples took only 49.3 milliseconds, compared to 19 minutes and 11 seconds for structured filters—a speedup of approximately 23,000x.

These results demonstrate that topological features, when enriched by random filters, provide a more robust and discriminatory starting point for deep learning models than raw pixels alone.

Conclusion and Future Directions

Random filters successfully bridge the gap between the stability of TDA and the discriminatory needs of complex classification tasks. While demonstrated here on 2D images, the concept of "interference topology" is broadly applicable. Future work will explore extending this framework to 3D medical imaging and point cloud data, as well as developing a more formal mathematical theory linking random filtering to topological stability.

Summary updated

Abstract

Topological representations of data are inherently coarse summaries which endows them with certain desirable properties like stability but also potentially inhibits their discriminatory power relat...
BibTeX
@inproceedings{jorgenson2022random,
  author = {Grayson Jorgenson and Henry Kvinge and Tegan Emerson and Colin C. Olson},
  title = {{Random Filters for Enriching the Discriminatory power of Topological Representations}},
  booktitle = {TAG-ML},
  year = {2022},
  url = {https://openreview.net/forum?id=c8W0zeJRLc6},
}
Figure