Fig. 4: Hardware properties of the G-KAN architecture.

a, d Topology of a KAN. Unlike MLPs, each edge is not merely a simple synapse but can be expanded into an activation function represented by a combination of Gaussian-like basis functions (Gaussian-like kernels), as illustrated in (d). b GMCs serve as the fundamental electrical units for performing tunable Gaussian-like basis functions, and are compatible with crossbar array architectures to realize analog in-memory VMM and MAC operations. c Schematic of the G-KAN. e Demonstration of function-level parallelism in a G-KAN during inference operations (m input nodes, p output nodes, with n basis functions between input-output nodes). f Basic performance evaluation of the G-KAN (with 100 basis functions) using a 1D function regression task as a representative example.