Wavelet Neural Network
Also known as: WNN, Wavelet-based neural network, Wavelet networks
A wavelet neural network (WNN) is a function approximation architecture that uses wavelet functions as activation functions in place of traditional sigmoid or ReLU functions. Introduced by Zhang and Benveniste (1992), WNNs combine the multiscale decomposition properties of wavelets with the learning capabilities of neural networks. The result is a flexible nonparametric model that can capture localized features and multi-resolution patterns efficiently, with fewer parameters and better interpretability than standard deep networks.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use WNN when you need interpretable, multiscale function approximation, especially for signals or images with features at multiple frequencies. They are ideal for nonlinear regression, time-series forecasting, classification with localized features, and denoising. WNN is preferable to deep networks when interpretability and sample efficiency matter.
Strengths & limitations
- Inherent multiscale decomposition—simultaneously captures features at multiple frequencies without explicit layer depth
- Good approximation properties with fewer parameters than standard MLPs, reducing overfitting and improving sample efficiency
- Interpretable hidden units with clear time-frequency localization, enabling feature visualization and understanding
- Fast convergence during training due to natural basis properties of wavelets
- Combines theoretical wavelet analysis with practical neural network learning
- Less developed software ecosystem compared to mainstream deep learning frameworks, limiting accessibility
- Approximation rate is slower than very deep networks for highly complex functions, requiring more hidden units
- Selection of wavelet type and initial scale ranges requires domain knowledge and experimentation
- Training procedure is less standardized than ReLU networks; optimization landscapes can be non-convex with multiple local minima
Frequently asked
How many hidden units (wavelets) do I need?
Start with 10–20 wavelets per input dimension and increase if needed. Use cross-validation to find the sweet spot between model complexity and generalization. More wavelets capture finer details but risk overfitting.
Which wavelet should I use?
Mexican hat and Morlet wavelets are common choices. Mexican hat is smoother and useful for approximation; Morlet is more oscillatory and better for frequency analysis. Try both and compare performance.
How do WNNs compare to convolutional neural networks (CNNs)?
CNNs are designed for hierarchical feature extraction from structured data (images). WNNs are more flexible and interpretable but less specialized for specific data types. For multiscale analysis with interpretability, WNN may be preferable; for image classification, CNN is usually better.
Can WNNs handle time-series classification?
Yes, though they require adapting the architecture for sequence modeling. You can use WNN as a feature extractor (learning wavelet bases) followed by classification layers, or extend it to recurrent variants.
Are trained wavelet parameters interpretable?
Yes. The learned scales tell you which frequencies the network found important. Learned shifts show spatial or temporal locations of features. Visualization of learned wavelets and their activations on sample data reveals what the network learned.
Sources
- Zhang, Q., & Benveniste, A. (1992). Wavelet networks. IEEE Transactions on Neural Networks, 3(6), 889–898. DOI: 10.1109/72.165591 ↗
- Pati, Y. C., & Krishnaprasad, P. S. (1992). Nonlinear dynamics and signal processing in the cochlea. ICASSP, pp. V373–V376. link ↗
- Misiti, M., Misiti, Y., Oppenheim, G., & Poggi, J. M. (1997). Wavelet Toolbox. The Mathworks. link ↗
How to cite this page
ScholarGate. (2026, June 3). Wavelet Neural Network. ScholarGate. https://scholargate.app/en/time-series/wavelet-neural-network
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Multilayer PerceptronDeep learning↔ compare
- Recurrent Neural NetworkDeep learning↔ compare