Less is More: Scaling Dynamics with Sparsity

Connecting every oscillator to every other oscillator makes Un-0 increasingly difficult to implement as it grows: doubling the oscillator count roughly quadruples the connections. In a follow-up to the original image-generator study, Unconventional AI explores deleting connections randomly or dividing oscillators into internally connected groups. The sparsity study tests both approaches.

The reported experiments found ranges where fewer connections improved model performance as well as reducing connectivity requirements. The authors associate this with avoiding excessive synchronization that limits useful dynamics. Removing too many connections eventually impairs communication and performance. Connectivity is therefore a design variable to tune, with the best setting depending on system size and structure.