Feasibility of Implementing KAN Layer Support via Extension API

[AbdessNassihi]

I'm exploring the possibility of integrating a Kolmogorov-Arnold Network (KAN) (https://pypi.org/project/pykan/) into an FPGA workflow using hls4ml, and I'm wondering whether it's feasible to implement this via the Extension API.
Do you foresee any limitations for this?

[calad0i]

The individual non-linear functions can be implemented as generalized lookup tables, and the aggregation is trivial, thus technically it is possible. Though, I do not see any practical application of this architecture to motivate the integration at the moment. Do you have a specific use case of it?

[AbdessNassihi]

Thanks for your reply! KANs promise similar performance to MLPs but with fewer parameters. This makes them an appealing candidate for hardware acceleration, and I plan to compare their resource utilization and inference efficiency. Still have to decide which application i will focus in.

[vloncar]

There are KANs in development for hls4ml, @Duchstf and @ag2718 can provide more information on the state.