KANMixer: a minimal KAN-centered mixer for long-term time series forecasting
The episode analyzes 'KANMixer,' a minimal model for long-term time series forecasting centered around Kolmogorov-Arnold Networks (KANs). Hosts discuss KAN's potential to simplify complex models, noting that B-spline bases and using KAN in the prediction head are key insights. Limitations include computational cost and inability to model cross-variable relationships.