A mathematical function that decomposes signals into localized time-frequency components at multiple scales.
A wavelet is a compact, oscillating mathematical function used to decompose signals or data into components across multiple scales of resolution. Unlike the Fourier transform, which represents signals as infinite sinusoidal waves and sacrifices temporal information for frequency precision, wavelets are localized in both time and frequency simultaneously. This dual localization is achieved through two operations: scaling (stretching or compressing the wavelet to capture different frequency bands) and translation (shifting it along the time axis to capture when those frequencies occur). The result is a rich, multi-resolution representation that captures both coarse global structure and fine local detail within the same framework.
The core computational mechanism is the wavelet transform, which comes in continuous and discrete forms. The discrete wavelet transform (DWT) is particularly practical for machine learning applications, recursively decomposing a signal into approximation coefficients (low-frequency content) and detail coefficients (high-frequency content) at successive levels. This hierarchical decomposition mirrors the way convolutional neural networks build representations across scales, and the two approaches share deep theoretical connections. Wavelet bases such as Haar, Daubechies, and Morlet offer different trade-offs between smoothness, compact support, and computational efficiency, allowing practitioners to choose a basis suited to their data's characteristics.
In machine learning, wavelets serve several important roles. They are widely used for feature extraction from time-series and audio signals, where non-stationary behavior — frequency content that changes over time — makes Fourier methods inadequate. Wavelet-based denoising, which thresholds detail coefficients to suppress noise while preserving signal structure, is a standard preprocessing step for physiological, financial, and sensor data. In computer vision, wavelet decompositions underpin classical image compression standards like JPEG 2000 and inform the design of modern architectures that process images at multiple resolutions.
Wavelets became relevant to machine learning in the 1990s as the discrete wavelet transform matured and computational resources made large-scale signal processing feasible. Their ability to compactly represent structured, non-stationary data with sparse coefficients remains valuable even as deep learning has absorbed many signal processing tasks, and wavelet-inspired ideas continue to influence neural network architecture design.