Rational Gaussian Wavelets and Corresponding Model Driven Neural Networks

Attila Miklós Ámon, Kristian Fenech, Péter Kovács, Tamás Dózsa

TL;DR

This paper introduces "Rational Gaussian Wavelets" (RGWs), a new class of adaptable wavelets whose shape can be precisely controlled by adjusting their poles and zeros. These RGWs, when integrated into a "Variable Projection" (VP) neural network architecture, enable highly interpretable and efficient feature learning for complex signals, outperforming existing methods in ECG arrhythmia detection.

What Problem Does This Paper Address?

Traditional wavelet transforms often rely on fixed "mother wavelets," which may not optimally capture the intricate details of complex signals. While adaptive wavelet methods exist, they can struggle with properties like translation invariance or smoothness, and their interpretability can be limited. This paper addresses the need for a wavelet transform that is both highly adaptable to signal morphology and interpretable, particularly for integration into machine learning models. The goal is to achieve sparse signal representations with fewer wavelet coefficients, leading to more efficient and understandable models, especially in critical applications like biomedical signal processing.

Methods

The core innovation is the introduction of Rational Gaussian Wavelets (RGWs). These are constructed by multiplying a Gaussian function with a rational function. The key idea is that the poles and zeros of this rational function become free parameters that can be tuned to precisely control the shape of the mother wavelet. This allows for the creation of wavelets that can mimic complex signal morphologies, including sharp peaks and non-symmetric shapes, while still maintaining desirable properties like smoothness and symmetry.

The paper also leverages Variable Projection (VP) techniques. VP operators are used to approximate wavelet coefficients efficiently. This approach is integrated into a neural network architecture called VP-Net. In this setup, the RGWs are used within a specialized "RGW-VP layer." This layer not only learns the optimal translation and scale parameters for the wavelets but also, crucially, learns the poles and zeros of the RGW mother wavelet itself. This end-to-end learning of wavelet morphology, combined with VP for efficient coefficient approximation, allows the network to learn highly relevant and interpretable features directly from the data.

Datasets

The paper uses the MIT-BIH arrhythmia database for training and testing. This dataset contains approximately 50,000 heartbeats and is noted as being unbalanced, reflecting real-world scenarios where normal heartbeats are more common than abnormal ones. The authors mention that many current approaches focus only on the QRS complex, but their work utilizes the entire heartbeat, including the P-wave, which is also relevant for recognizing certain arrhythmias.

Key Insights and Ideas

  • Tunable Wavelet Morphology: The ability to precisely control wavelet shape by adjusting poles and zeros is a significant advancement. This allows wavelets to be tailored to specific signal characteristics, leading to more efficient representations.
  • Interpretable Feature Learning: By learning the wavelet parameters (poles and zeros) directly within a neural network, the model can generate features that correspond to medically relevant signal segments (e.g., P-wave, QRS complex). This makes the learned features more interpretable for domain experts.
  • Sparse Representations: The adaptive nature of RGWs, combined with VP, allows for the approximation of signals using a significantly smaller number of wavelet coefficients compared to traditional methods. This leads to more compact and computationally efficient models.
  • End-to-End Optimization: The RGW-VP layer enables the simultaneous optimization of wavelet morphology and feature extraction parameters within a single neural network, driven by gradient-based methods.

Key Results

The proposed RGW-VP-Net architecture demonstrated superior performance in classifying ventricular ectopic beats (VEBs) from ECG signals compared to other methods, including VP-Nets using Ricker wavelets or adaptive Hermite functions. Specifically, the RGW-VP model achieved higher accuracy, sensitivity, and positive predictivity across the board. For instance, it outperformed the Ricker-VP in the VEB +P metric by approximately 10%. The paper also highlights that the learned RGW wavelets and their corresponding coefficients were interpretable, aligning with cardiologists' criteria for VEB detection.

Limitations and Future Work

  • Computational Cost: While the RGW-VP approach leads to sparse representations, the initial computation of wavelet coefficients and their derivatives can be computationally intensive. The paper mentions that the admissibility proof holds for a broader class of rational functions (W), suggesting potential avenues for further exploration.
  • Generalizability: The study primarily focuses on ECG classification. The authors suggest exploring other applications for RGW-VP, such as industrial fault detection and univariate function layers in neural networks.
  • Parameter Tuning: While the RGW-VP architecture is designed for gradient-based optimization, the initial manual grid search for hyperparameters in the experiments indicates that further research into automated hyperparameter optimization could be beneficial.

Summary updated

Abstract

In this paper we introduce a highly adaptive continuous wavelet transform using Gaussian wavelets multiplied by an appropriate rational term. The zeros and poles of this rational modifier act as free parameters and their choice highly influences the shape of the mother wavelet. This allows the proposed construction to approximate signals with complex morphology using only a few wavelet coefficients. We show that the proposed rational Gaussian wavelets are admissible and provide numerical approximations of the wavelet coefficients using variable projection operators. In addition, we show how the proposed variable projection based rational Gaussian wavelet transform can be used in neural networks to obtain a highly interpretable feature learning layer. We demonstrate the effectiveness of the proposed scheme through a number of numerical experiments including biomedical applications, and the detection of abnormal road surface based on tire sensor signals.
BibTeX
@article{amon2025rational,
  author = {Attila Miklós Ámon and Kristian Fenech and Péter Kovács and Tamás Dózsa},
  title = {{Rational Gaussian Wavelets and Corresponding Model Driven Neural Networks}},
  journal = {IEEE Trans. Signal Process.},
  year = {2025},
  url = {https://openreview.net/forum?id=ptXCMhZ9nw},
}
Figure