Skip to content

Mapping KAN Latents to Features

Home / Open Questions / Mapping KAN Latents to Features

Background: In Kolmogorov–Arnold Network (KAN) architectures, each edge represents a learnable 1D function over latent coordinates, allowing for direct visualization of learned nonlinearities. However, these latent representations are computed on transformed patch embeddings rather than raw input variables.

Question / Future Work: There is a need for robust attribution methods that can map the visually inspectable 1D nonlinearities learned by KAN edge functions back to raw, human-interpretable input variables. While the functional form itself is inspectable, its current representation in latent space obscures the direct mechanistic interpretation of how raw physical variables interact within the model.

Why It Matters: Bridging the gap between ‘latent-space inspectability’ and ‘feature-space interpretability’ is critical for scientific validity, ensuring that identified patterns correspond to physical phenomena in the raw time series.

Metadata & Links