MAML initialization accelerates distributed active noise control convergence

Researchers from Nanyang Technological University and collaborators have introduced a meta-learning approach to solve a long-standing bottleneck in distributed multichannel active noise control (DMCANC). In DMCANC, multiple nodes run local single-channel ANC controllers and exchange information to cancel noise over large areas. But these systems typically start from zero or random filter coefficients, forcing adaptive filters to slowly converge and hampering inter-node collaboration. The team's solution, described in arXiv paper 2607.29117, applies model-agnostic meta-learning (MAML) to learn an optimal initialization that generalizes across heterogeneous acoustic environments, including variations in primary and secondary paths.

Using numerical simulations with broadband and real-world noise recordings, the proposed MAML-initialized DMCANC achieved substantially faster convergence than conventional zero/random initialization, especially under stationary and time-varying noise conditions. The learned init also improved steady-state noise reduction, demonstrating that meta-learning can encode cross-node acoustic knowledge into a reusable starting point. This work builds on the same MAML framework popularized in few-shot learning, repurposing it for adaptive filter initialization in distributed signal processing. The results highlight MAML as a promising tool for scaling ANC to larger spaces—such as open-plan offices, smart factories, and urban infrastructure—where fast, coordinated noise suppression is critical but current systems take too long to adapt.