CuMPerLay: Learning Cubical Multiparameter Persistence Vectorizations

AI summary

Overview

Research area: Topological data analysis (TDA) applied to computer vision, specifically differentiable multiparameter persistence vectorization for deep learning pipelines.

Technical level: Advanced. The paper builds up cubical complexes, cubical (multi)parameter persistence, and vectorization theory from first principles, but it assumes familiarity with homology, persistence diagrams, and modern vision architectures.

Scope: The paper introduces a differentiable layer that turns cubical multiparameter persistence into learnable feature vectors and embeds it inside Swin Transformer–based networks for medical image classification and semantic segmentation.

What This Paper Is About

Deep networks extract pixel-wise features but often miss the global geometric and topological structure of images. Topological data analysis, and in particular persistent homology, can describe connectivity and shape, but the richer multiparameter persistence (MP) variant has been hard to use on images because there is no natural way to vectorize the output of a multifiltration over the cubical complexes that images natively form. The goal of this paper is to make cubical multiparameter persistence differentiable and learnable so that it can be dropped into standard deep learning architectures and trained end-to-end.

Key Contributions

1. A differentiable CuMPerLay vectorization. The authors propose a novel vectorization for Cubical Multiparameter Persistence that combines learnable filtration functions with learnable vectorizations and can be integrated into state-of-the-art image-based learning pipelines.
2. Theoretical stability guarantees. They prove the stability of CuMPerLay, along with a family of other filtration functions they devise, under generalized Wasserstein metrics — establishing that CuMPerLay is stable as a corollary.
3. A hybrid topology-aware deep model (TopoSwin). They build a model integrating cubical multiparameter persistence with Swin transformers, reporting synergistic improvements and benefits on classification and segmentation, particularly in limited-data scenarios.
4. The first public CUDA implementation. To the authors' knowledge, this is the first publicly available CUDA GPU implementation of Cubical Persistence and Cubical Multiparameter Persistence (prior MP tooling such as multipers is CPU-only for simplicial complexes).

The source code is released at https://github.com/circle-group/cumperlay.

Main Findings

• Multiparameter persistence needs a vectorization workaround: Because of the partial ordering problem, a topological feature's region of persistence in a multifiltration need not resemble a rectangle, so there is no natural extension of persistence diagrams (or their vectorization) to MP outputs. CuMPerLay addresses this by decomposing the CMP into a combination of individual, learnable single-parameter filtrations.
• Slicing plus stability propagates: Applying horizontal slices to an M×N multipersistence grid and vectorizing each slice gives M vectors. If the underlying single-persistence vectorization is stable, the induced multipersistence vectorization is stable as well (Theorem 1). Since PersLay vectorizations are stable, CuMPerLay itself is stable (Corollary 1), with the bound D(Ψ_MP(K_F), Ψ_MP(K_G)) ≤ Ĉ_{Ψ_MP} ‖F − G‖_∞.
• Improved stability claim over SP counterparts: The theoretical analysis is stated to show improved stability when compared to single-parameter counterparts.
• Empirical benefits on benchmarks: Experiments on benchmark medical imaging and computer vision datasets show that the topological vectors improve model performance in both standard and limited-data scenarios for classification and segmentation. Specific numerical results, benchmark scores, and comparisons are not reported in the content provided; the paper's experiments section is truncated before any result tables or metric values.
• Learnable filtrations change with depth: The paper reports sample compact multifiltration representations and corresponding bifiltrations learned from the third layer of the network on the ISIC dataset (shown in their Figure 5), with further examples promised in supplementary material.
• Topological baselines: Standalone topological features were also evaluated without the deep model — SP vectors via 100-bin Betti curves per homology group, and MP vectors from threshold arrays (color: 10 thresholds per channel giving a 10×10×10 representation per homology group; grayscale: 50 thresholds with 10 erosion filtration levels producing a (10, 50) array per homology group), flattened with the top 300 features selected via XGBoost feature importance and fed to an MLP with two ReLU hidden layers and 0.3 dropout.

Methodology in Plain English

The core idea. Images are naturally cubical complexes — grids of vertices, edges, squares, and voxels. Tracking how topological holes appear and disappear across two or more varying thresholds (a multifiltration) gives a much finer description than a single threshold sequence. The obstacle is that the output of a multifiltration is not a neat list of birth–death pairs, so it cannot be fed to a neural network directly.

The workaround: slicing. The authors place an M×N grid over the two-parameter filtration and take horizontal slices. Each slice is an ordinary single-parameter filtration whose persistence diagram can be vectorized with a standard, well-understood method. This yields M separate vectors, one per row.

Making it learnable. Each slice is vectorized with a level-dependent PersLay-style scheme: a sum aggregation, a power weighting function |d − b|^{w_t} whose powers are learnable, and a triangle point transformation with learnable sample times. The resulting vectors are combined by an aggregator ρ, modeled as a multi-layer perceptron. Because the whole pipeline is differentiable, gradients can be pushed back through the persistence computation via straight-through gradients into the pixels that generated the persistence pairs, and even into the filtration functions themselves, which are modeled as neural networks.

The network: TopoSwin. Four components are stacked: (i) a multifiltration decoder, (ii) a differentiable GPU cubical multipersistence block, (iii) the CuMPerLay vectorization module, and (iv) a fusion block. The decoder upsamples input images or intermediate activations and outputs M channels; a staircase function with straight-through gradients quantizes each row into N+1 levels, enforcing a valid filtration within each row, while a soft regularization loss enforces the column structure needed for a genuine bifiltration. A residual gated fusion block, FB(X, v) = X + X ⊗ σ(W₁v) + W₂v, recalibrates channel activations using the vectorization. Training adds a topological classification head (two-layer MLP with cross-entropy) and a negative local image entropy loss on the compact filtration outputs alongside the standard classification loss.

Evaluation. Four public datasets are used: ISIC 2018 (10,015 dermoscopic images across seven diagnostic categories, official split), CBIS-DDSM (3,568 mammograms; 2,111 benign, 1,457 malignant; official train/test split with an 80:20 train/validation split), a glaucoma dataset drawn from the Eye Disease Image Dataset (1,024 healthy and 1,349 glaucoma fundus images; originals of 2004×1690 resized to 224×224; 80:20 train/test with an 80:20 validation split of the training portion), and PASCAL VOC 2012 with the Semantic Boundaries augmentation (excluding VOC validation images; 90:10 split of the 7,087 training images for validation; VOC validation split of 1,449 images as test; all resized to 224×224).

Model configuration. The base is SwinV2-B with ImageNet-1K pretrained weights, with one learnable cubical multipersistence module after each Swin transformer block plus one on the input images. Filtration uses 8×16 MP grids and 112×112 filtration image size, except the input-level module which uses 224×224. Vectorization uses 100 learnable sample locations. Training uses AdamW at a learning rate of 10⁻⁴, weight decay 0.01, no learning rate scheduling, random flip augmentation only, and 100 epochs; loss weights are 1.0 for cross-entropy, 0.25 for the topological loss, and 0.01 for the regularization loss (the local entropy weight is given as 10⁻⁴ in a sentence truncated in the provided text). The single-persistence variant is denoted TopoSwin-SP (TS-SP) and the multipersistence variant TopoSwin-MP (TS-MP).

Why This Matters

Research impact. Multiparameter persistence has been held back in machine learning by theoretical obstacles (the partially ordered threshold set) and by the lack of vectorization suited to cubical complexes. This work supplies both a differentiable vectorization and stability guarantees, plus a released CUDA implementation, which lowers the barrier for other researchers to use MP in vision pipelines. It also extends MP analysis beyond the specific filtrations and Betti numbers studied in earlier topology-aware image work.

Real-world applications.

• Medical image classification and diagnosis — the experiments target dermoscopic melanoma classification, mammogram-based breast cancer diagnosis, and glaucoma detection from fundus images, where limited labeled data is common.
• Limited-data clinical settings — the paper specifically highlights benefits in limited-data scenarios, which matters where annotation is expensive or rare.
• Semantic segmentation — evaluated on PASCAL VOC 2012 with Semantic Boundaries augmentation, relevant to scene understanding and structured image parsing.
• Topology-aware image processing — prior MP-based work on mathematical morphology and denoising indicates possible uses in restoring structure-preserving image content.

Industry relevance. Topological features can be computed as an auxiliary module bolted onto existing backbones such as Swin transformers, so the approach is compatible with current pretrained model workflows. Domains where global shape and connectivity matter more than local texture — medical imaging, materials inspection, and structured image analysis — are natural early adopters, and the GPU implementation addresses the practical bottleneck that kept MP mostly out of deep learning.

Future Directions

• Learning better bifiltrations. The paper states that determining an optimal bifiltration is a complex challenge and an active research area; the data-driven multifiltration decoder is one attempt, leaving room for improved learned filtration designs.
• Scaling to higher dimensions. The construction is stated to extend naturally beyond 2D images, for example to 3D voxel grids, which is untested here.
• Beyond slicing. Slicing is used because MP outputs such as persistence diagrams are technically hard to obtain; alternatives such as higher-dimensional persistence landscapes or signed-barcode/Radon-measure vectorizations avoid slicing-related information loss and direction dependence, and generalizing them to cubical complexes remains open.
• Refining topological priors and losses. The authors' related-work discussion points to refining topological priors, differentiable persistence computations, and domain-specific loss functions as open directions, along with more thorough evaluation across diverse filtrations and tasks.

Target Audience

Researchers and practitioners in topological data analysis and computer vision who want to incorporate multiparameter persistence into deep networks; medical imaging and computational pathology researchers working with limited labeled data; machine learning engineers interested in hybrid architectural priors beyond convolutional features; and readers seeking a bottom-up exposition of cubical complexes and cubical multipersistence aimed explicitly at the computer vision community.

Authors’ abstract

We present CuMPerLay, a novel differentiable vectorization layer that enables the integration of Cubical Multiparameter Persistence (CMP) into deep learning pipelines. While CMP presents a natural and powerful way to topologically work with images, its use is hindered by the complexity of multifiltration structures as well as the vectorization of CMP. In face of these challenges, we introduce a new algorithm for vectorizing MP homologies of cubical complexes. Our CuMPerLay decomposes the CMP into a combination of individual, learnable single-parameter persistence, where the bifiltration functions are jointly learned. Thanks to the differentiability, its robust topological feature vectors can be seamlessly used within state-of-the-art architectures such as Swin Transformers. We establish theoretical guarantees for the stability of our vectorization under generalized Wasserstein metrics. Our experiments on benchmark medical imaging and computer vision datasets show the benefit CuMPerLay on classification and segmentation performance, particularly in limited-data scenarios. Overall, CuMPerLay offers a promising direction for integrating global structural information into deep networks for structured image analysis.