FPGA-Based Adaptive LIF Neuron for High-Speed, Energy-Efficient Spiking Neural Network

Spiking Neural Networks (SNNs) have drawn attention due to their biological behaviour and low energy consumption compared to conventional neural networks. The efficiency of SNN depends on the neurons, and thus, proper selection of neurons becomes imperative for exhibiting appropriate biological behaviour. This paper realizes an adaptive <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Leaky Integrate-and-Fire</i> (LIF) neuron, modeled using presynaptic spike inputs, adaptive threshold, and adaptive leaky rate. The proposed LIF neuron employs a low firing rate, which reduces the computational cost and floating-point operations at different presynaptic input frequencies and noise levels. Comparatively, the proposed LIF neuron significantly improves resource utilization and latency. The number of flipflops, Look-Up Table (LUT), and latency of this LIF neuron are respectively 2.5×, 5.7×, and 2.0× better than the best-performing contemporary LIF neuron. To evaluate the performance of proposed neuron on the network, we implement an SNN on an Artix-7 Field Programmable Gate Array (FPGA). This SNN achieves 95% accuracy on the Modified National Institute of Standards and Technology (MNIST) handwritten dataset classification. The total delay in classifying an image using our proposed neuron-based SNN is 0.9 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i>, 60% less than the best-performing contemporary SNN. This SNN is 68% more energy-efficient in image processing.