What is behind the meta-learning initialization of adaptive filter? - A naive method for accelerating convergence of adaptive multichannel active noise control

Abstract

Active noise control (ANC) is a typical signal-processing technique that has recently been utilized extensively to combat the urban noise problem. Although numerous advanced adaptive algorithms have been devised to enhance noise reduction performance, few of them have been implemented in actual ANC products due to their high computational complexity and slow convergence. With the rapid development of deep learning technology, Meta-learning-based initialization appears to become an efficient and cost-effective method for accelerating the convergence of adaptive algorithms. However, few dedicated Meta-learning algorithms exist for adaptive signal processing applications, particularly multichannel active noise control (MCANC). Hence, we proposed a modified Model-Agnostic Meta-Learning (MAML) initialization for the MCANC system.¹ Additional theatrical research reveals that the nature of MAML, when applied to signal processing, is the expectation of a weight-sum gradient. Based on this discovery, we devised the Monte-Carlo Gradient Meta-learning (MCGM) algorithm, which employed a more straightforward procedure to accomplish the same performance as the Modified MAML algorithm. Furthermore, the numerical simulation of ANC using raw noise samples on measured paths validates the efficacy of the proposed methods in accelerating the convergence of the multichannel-filtered reference least mean square algorithm (McFxLMS).

Keywords: Active noise control; Filtered reference least mean square algorithm; Meta-learning.