Neural Network-Based Signal Processing in VLSI Circuits

This work investigates the integration of neural networks (NN) in the levels of very-large-scale integration (VLSI) to improve signal processing. We proposed and realized a neural-network-based VLSI based on NNs, and we compared it with a conventional DSP in terms of SNR, processing time (latency), energy consumption, and accuracy of signal processing tasks. Our results demonstrate that the NN-based circuit effectively increases SNR by up 7 dB compared to DSP methods. Latency measurements are obtained showing that the NN-based circuit has average latencies of $\mathbf{1. 2}$ microsecond and 2.8 microsecond for single and complex tasks, respectively, better than the DSP's latencies of $\mathbf{3. 5}$ microsecond and $\mathbf{6. 2}$ microsecond. Furthermore, the NN implementation decreases the average power by 40 % with 45 mW power consumption whereas DSP consumes 75 mW. Accuracy analysis shows that NN-based circuits outperforms in signal-processing operations with accuracies of 98.2 % for signal filtering, 97.5 % for denoising, and 99.1% for pattern recognition than DSP 95.7%, 92.3 % and 96.8 % respectively. These findings emphasize that the NN-based circuit can offer superior precision, throughput, and energy efficiency. To sum up, the implementation of neural networks on VLSI can provide significant advantages compared with DSP-based implementation, so that it appears as a promising technology for advanced signal processing that requires high performance, high efficiency, and high precision. This study establishes the basis for future progress of VLSIutilization of NN-based circuits in generic signal processing applications such as communications, imaging, and real-time data analysis.