Kolmogorov-Arnold Autoencoder for Representation Learning
An autoencoder built from Kolmogorov-Arnold Network layers instead of multi-layer perceptrons: each layer applies learnable univariate functions on its edges rather than a fixed activation on a linear map. The learner sees how replacing MLP layers with KAN layers that use learnable polynomial functions (in place of the usual B-splines) sharpens the latent representation, and why this improves reconstruction quality and downstream retrieval, classification, and denoising.
2501.00420
The Kolmogorov-Arnold Auto-Encoder (KAE) is an autoencoder whose encoder and decoder layers are Kolmogorov-Arnold Network (KAN) layers rather than ordinary multi-layer-perceptron layers. Where an MLP…