Abstract
Spiking Neural Networks (SNNs) emulate the computational prowess and energy efficiency of the human brain. However, deploying SNNs practically can often pose challenges due to hardware constraints. This paper introduces a method that effectively tackles this problem through hardware-aware network selection and quantization of SNNs, thus bridging the gap between neural network architectures and hardware capabilities. The effectiveness of this approach is tested on three diverse datasets: Fashion MNIST (FMNIST), SHD-10, and a QT database Electrocardiogram (ECG) data set. Notably, our method achieves competitive quantized accuracies of 85.7%, 85.7%, and 85.19% on these datasets respectively. These results are significant as they are achieved with the use of qint8 precision, demonstrating only a minor accuracy loss from the full-precision float32 counterparts, despite significant reductions in model complexity to meet hardware constraints. Additionally, we propose a mixed-signal implementation of the Leaky Integrate-and-Fire (LIF) neuron, taking advantage of the benefits of both domains and making it compatible with In-Memory Computing (IMC) accelerators. By leveraging the benefits of non-volatile memory technologies, this research facilitates the deployment of SNNs on real-world hardware accelerators with minimal accuracy loss. Our work is instrumental in highlighting the potential of mixed-signal IMC in balancing flexibility and power efficiency trade-offs, making it particularly valuable for ultra-low power edge devices.