2026A | Controlling Frequency Learning through Wavelet-Based Kolmogorov–Arnold Networks скачать в хорошем качестве

2026A | Controlling Frequency Learning through Wavelet-Based Kolmogorov–Arnold Networks

Speaker: Juan Daniel Meshir Vargas Abstract: Artificial neural networks tend to learn low-frequency components of a target function faster than high-frequency components, often failing to capture fine-scale details. This phenomenon, known as spectral bias, limits their performance in applications involving oscillatory signals, sharp transitions, or high-frequency solutions of partial differential equations. Recently, Kolmogorov–Arnold Networks (KANs), inspired by the Kolmogorov–Arnold representation theorem, have emerged as a promising alternative to classical multilayer perceptrons. In this talk, we present a wavelet-based extension of KANs (Wav-KANs) that enables explicit control over frequency learning through the frequency parameter of the mother wavelet. We show how this mechanism mitigates spectral bias and improves the learning of high-frequency components. Theoretical insights and connections to the neural tangent kernel perspective will be discussed, highlighting the role of architectural design in shaping frequency dynamics.