Topological Data Analysis (TDA) provides a unique lens for understanding data by focusing on its "shape"—identifying features like connected components, loops, and voids that persist across different scales. The primary output of this analysis is the Persistence Diagram (PD): a multiset of points in the 2D plane where each point $(b, d)$ represents the "birth" and "death" of a topological feature.
While PDs are mathematically elegant and stable, they are notoriously difficult to use in machine learning. Standard algorithms like Support Vector Machines or Neural Networks expect fixed-length vectors or tensors, but PDs are variable-sized sets. Bridging this gap has evolved from rigid mathematical kernels to flexible, learnable vectorization layers.
The Impedance Mismatch of Persistence Diagrams
A persistence diagram is essentially a summary of how long different "holes" in a dataset survive as you thicken the data points. Points far from the diagonal ($d=b$) represent robust features, while points near the diagonal are often considered noise.
The challenge is that the space of persistence diagrams does not naturally possess a linear structure. You cannot simply "average" two diagrams or feed them into a standard loss function. Early attempts to solve this relied on kernel methods. Jan Reininghaus and colleagues introduced A stable multi-scale kernel for topological machine learning, which treated PD points as heat sources and compared diagrams by looking at the resulting heat distributions. Similarly, Genki Kusano and Yasuaki Hiraoka developed the Persistence weighted Gaussian kernel to weight features by their significance.
However, these kernels suffer from a "quadratic bottleneck." Comparing $N$ datasets requires calculating an $N \times N$ similarity matrix, where each entry involves comparing two sets of points. For large-scale datasets, this $O(N^2)$ scaling makes topological analysis computationally prohibitive.
From Kernels to Fixed Vectors
To achieve scalability, researchers shifted toward vectorization: transforming a PD into a fixed-length numerical vector before it reaches the machine learning model.
One intuitive approach is inspired by computer vision: the "Bag-of-Words" model. In Persistence Bag-of-Words for Topological Data Analysis, Bartosz Michał Zieliński and Pawel Dlotko proposed partitioning the birth-death plane into regions. By counting how many topological features fall into each region, a complex diagram is reduced to a simple histogram. This allows TDA to scale to massive datasets because each diagram is processed independently into a vector.
Building on this, Martin Royer and Frédéric Chazal introduced ATOL: Automatic Topologically-Oriented Learning. ATOL uses unsupervised quantization (like k-means) to automatically find the most "informative" regions of the birth-death plane across a collection of diagrams. This "spatial" approach ensures the vectorization focuses on where the data actually lives, rather than using a rigid, inefficient grid.
Learnable Topological Representations
The most recent evolution treats vectorization not as a preprocessing step, but as a part of the neural network itself. Instead of humans deciding how to bin the diagrams, the model learns which topological features matter for a specific task.
Christoph D. Hofer and Roland Kwitt pioneered this in Deep Learning with Topological Signatures. They designed a neural network layer that applies a parameterized function to each point in the PD and then aggregates the results. Because this process is differentiable, the network can use backpropagation to "move" its sensors to the most discriminative parts of the topological space.
This idea culminated in PersLay, a unified framework developed by Mathieu Carrière and Théo Lacombe. In A General Neural Network Architecture for Persistence Diagrams and Graph Classification, they showed that almost all previous vectorization methods (like persistence images or landscapes) are special cases of a single permutation-invariant layer.
By treating the persistence diagram as a structured input to a deep model, PersLay allows the network to learn complex, task-specific topological features. This moves TDA from a niche mathematical tool to a standard component of the deep learning toolbox, capable of handling large-scale graph classification and 3D shape analysis with the same efficiency as traditional features.