Search papers, labs, and topics across Lattice.
This paper introduces HYDRA, a hyperbolic extension of Kolmogorov-Arnold Networks (KANs) that addresses parameter redundancy by utilizing spline-based functional learning within a bounded hyperbolic latent space. By mapping vector-valued inputs into the Poincar茅 ball and employing low-rank prototype blocks for shared functional transformations, HYDRA enhances both parameter efficiency and interpretability of representations. Extensive experiments show that HYDRA not only maintains competitive predictive performance across eight benchmark datasets but also improves training stability and reduces redundancy in model parameters.
Reducing parameter redundancy in KANs, HYDRA achieves superior predictive performance while enhancing interpretability in hyperbolic spaces.
Kolmogorov-Arnold Networks (KANs) enhance nonlinear function approximation by replacing scalar weights with learnable univariate functions. However, assigning an independent function to every connection results in substantial parameter redundancy, limiting their scalability and efficiency. To reduce this redundancy, we introduce \textbf{HY}perbolic \textbf{D}ynamic \textbf{R}epresentation \textbf{A}rchitecture (HYDRA), a parameter-efficient hyperbolic extension of KAN that combines spline-based functional learning with representations in the Poincar茅 ball. HYDRA maps vector-valued inputs into a bounded hyperbolic latent space, performs KAN-style updates in tangent space, and employs a low-rank prototype block to share functional transformations across hidden dimensions. The resulting hyperbolic representations provide a structured radial coordinate for interpretation, while radius control improves training stability by preventing boundary saturation. Extensive experiments across eight benchmark datasets demonstrate that HYDRA consistently achieves competitive or superior predictive performance while improving parameter efficiency and representation interpretability.
