This paper presents a power-efficient and re-configurable radial basis function-artificial neural network (RBF-ANN) for real-time pattern classification tasks. We have developed a MATLAB-based behavioral model, which can be employed to facilitate the off-chip learning of the proposed network with a hybrid learning algorithm. Besides, a generic design methodology is presented to ease implementing different scales of neural networks. For the RBF-ANN, an analog tunable Gaussian kernel circuit is used as an activation neuron while a current-mode computation is employed to improve the power efficiency. The proposed network is designed and simulated in 0.18 μm X-FAB CMOS process for a non-linear XOR-pattern classification and voice activity detector (VAD), respectively. It is found that the network's weights obtained using the developed model in MATLAB are well-matched with the Spectre simulation (Virtuoso) results with a maximum relative error of 7.1%. Consequently, the proposed model can be effectively employed for the off-chip training of RBF-ANN. Besides, the preliminary simulation results of the VAD depict that the classification accuracy reaches as large as 95% in a noisy environment of babble with an acoustic signal to noise ratio (SNR) of 10 dB. Meanwhile, the power efficiency of the proposed network gets significantly improved compared to the recent prior arts.

本文提出了一种用于实时模式分类任务的高效节能且可重新配置的径向基函数人工神经网络（RBF-ANN）。我们已经开发了基于MATLAB的行为模型，该模型可用于通过混合学习算法促进拟议网络的片外学习。此外，提出了一种通用的设计方法，以简化实现不同规模的神经网络的工作。对于RBF-ANN，将模拟可调谐高斯核电路用作激活神经元，同时采用电流模式计算来提高功率效率。拟议网络的设计和仿真精度为 0.18μ分别用于非线性XOR模式分类和语音活动检测器（VAD）的X-FAB CMOS工艺。结果发现，使用MATLAB中开发的模型获得的网络权重与Spectre仿真（Virtuoso）结果完全匹配，最大相对误差为7.1％。因此，所提出的模型可以有效地用于RBF-ANN的芯片外训练。此外，VAD的初步仿真结果表明，在噪声为10 dB的喧闹喧闹的嘈杂环境中，分类精度高达95％。同时，与最近的现有技术相比，所提议的网络的功率效率得到了显着改善。