A compact neuromorphic architecture with dynamic multiplexing to efficiently compute a nearest Kronecker product decomposition based RLS-NLMS algorithm for active noise control headphones

Abstract

Recently, embedded applications in resource-constrained electronic appliances are increasingly being used for noise reduction. Specifically, modern headphones are used to mitigate the environmental noise effects using advanced active noise control (ANC) systems. Despite achieving great performance, there is still a great challenge to develop compact ANC headphones since these devices contain a limited area. In addition, there are more challenges linked to improvement of the convergence rate, tracking, and complexity of the adaptive algorithm which is used in these devices to efficiently reduce the environmental noise. Specifically, the convergence properties can be improved by using the cutting-edge advanced adaptive algorithms along with the nearest Kronecker product (NKP) decomposition. In this work, we present a new neuromorphic architecture to efficiently compute an improved variant of the Kronecker product recursive least-squares (RLS). To achieve this architecture, we present three contributions; (1) we made a combination between the filtered-X RLS algorithm and normalized least mean squares (NLMS) algorithm to decrease the computational complexity by reducing the number of arithmetic operations required to efficiently compute the filter coefficients. Therefore, the proposed method requires fewer operations when compared with conventional RLS and RLS-NKP algorithms, respectively; (2) we use the spiking neural P (SN P) systems along with their advanced variants, such as rules on the synapses, colored spikes and dendritic delays to design two compact parallel arithmetic circuits (adder and divisor). In this way, the proposed method can be computed at high processing speeds and expending low area; (3) we propose a new digital neuromorphic architecture to be applied in active noise cancellation in headphones. Specifically, we propose a new adaptive unit core, which contains the proposed parallel neural arithmetic circuits, to perform dual filter operations, i.e, the proposed unit is capable of simulating two adaptive algorithms by using the same core. To achieve this we use the dynamic multiplexing technique. Therefore, our proposal exhibits low area consumption. As a consequence, the implementation of the proposed method can be easily integrated into resource-constrained ANC headphones.