Abstract: Recent years have witnessed an increased interest in the application of persistent homology, a topological tool for data analysis, to machine learning problems. Persistent homology is known for its ability to numerically characterize the shapes of spaces induced by features or functions. On the other hand, deep neural networks have been shown effective in various tasks. To our best knowledge, however, existing neural network models seldom exploit shape information. In this paper, we investigate a way to use persistent homology in the framework of deep neural networks. Specifically, we propose to embed the so-called "persistence landscape," a rather new topological summary for data, into a convolutional neural network (CNN) for dealing with audio signals. Our evaluation on automatic music tagging, a multi-label classification task, shows that the resulting persistent convolutional neural network (PCNN) model can perform significantly better than state-of-the-art models in prediction accuracy. We also discuss the intuition behind the design of the proposed model, and offer insights into the features that it learns.
Introduction
This research represents a pioneering effort to bridge the gap between Topological Data Analysis (TDA) and Deep Learning in the context of audio processing. By capturing the "shape" and structural stability of audio features, the proposed model significantly outperforms standard CNNs in automatic music tagging tasks.
The Problem: Capturing Topological Shape
While Convolutional Neural Networks (CNNs) are excellent at detecting local patterns and Recurrent Neural Networks (RNNs) excel at temporal dependencies, they often struggle to explicitly characterize the "topological shape" of a signal.
In audio, features like rhythm, timbre, and instrument presence have structural properties that persist across different scales. Standard neural networks lack a dedicated mechanism to measure how "stable" or "significant" a peak in a feature map is relative to the noise around it. This paper addresses this gap using Persistent Homology, a tool from TDA that numerically describes the birth and death of structural features (like connected components or peaks) as a threshold varies.
Methods: The PCNN Architecture
The core innovation is the integration of a Persistence Layer into a standard CNN pipeline, creating the Persistent Convolutional Neural Network (PCNN).
- Cubical Complexes: 1D audio feature sequences are treated as "cubical complexes." The authors use a superlevel-set filtration, analyzing the signal from highest to lowest values to observe how "connected components" (peaks) appear and merge.
- Persistence Landscapes: Raw topological data (birth/death pairs) is converted into Persistence Landscapes ($\lambda_k$). Unlike traditional persistence diagrams, landscapes are stable functions that can be sampled into matrices, making them compatible with standard backpropagation and matrix operations.
- Architecture Integration: As shown in the architecture diagram, the model processes feature maps through parallel paths: a standard middle convolution and a persistence layer. These features are then concatenated for final prediction.
Datasets and Evaluation
The model was evaluated on the MagnaTagATune dataset, a benchmark for music auto-tagging consisting of 25,863 clips (29 seconds each) with 188 tags covering instruments, genres, and moods. The evaluation primarily focused on the top 50 tags.
Key Insights
- Topological Stability: Persistence landscapes are inherently stable; small perturbations or noise in the audio signal result in only minor changes to the landscape, enhancing model robustness.
- Physical Interpretation: The authors discovered that persistence landscapes are highly correlated (Pearson’s $r \approx 0.97$) with onset strength, suggesting the TDA layer naturally learns to act as a sophisticated rhythm and beat detector.
- Complementary Information: The persistence layer captures structural relationships between peaks that standard convolutions might miss. Higher-order landscape pieces (higher $k$) were found particularly useful for "gentler" music like classical or choral, capturing subtle signal fluctuations.
Results
The PCNN demonstrated a clear performance boost over both pure CNNs and pure Persistent Neural Networks (PNNs).
As detailed in the results table, the PCNN (with $P=5$ landscape pieces) achieved a per-class AUC of 0.9013, outperforming the then-state-of-the-art models (0.898). Notably, even when a standard CNN was expanded to 6,400 filters to match the PCNN's dimensionality, it could not match the PCNN's performance, confirming that topological features provide unique, non-redundant information.
Limitations and Future Work
While successful, the study focused on 0-dimensional homology (connected components). Future research could explore 1-dimensional homology (loops or holes) to capture complex cyclical patterns, though this is computationally more demanding. Additionally, while the persistence layer was placed after early convolutions to manage overhead, making landscape sampling parameters learnable remains a promising avenue for future optimization.
Summary updated