Resonate-and-Fire Neurons as Frequency Selective Input Encoders for Spiking Neural Networks

More and more embedded applications demand to cope with complex data while still being energy-efficient. Neural networks provide the processing capabilities, but often cannot be utilised because of restricted power goals. Spiking neural networks have been shown to potentially solve this problem due to their hardware friendliness and energy efficiency. One remaining problem is the conversion of input data into event-based spikes in order to be processed. In this study, we examine using resonating neuron models to perform spectral transform and temporal spike encoding simultaneously directly on the analogue signal. Additionally, we compare the approach with the fast Fourier transform and demonstrate that both show comparable results while the neuromorphic realisation consumes significantly less energy. With the resonating neurons as input stage for large spiking neural networks, it is possible to realise energy efficient networks in neuromorphic hardware without the need of any digital logic. A limitation of the approach is the comparatively large silicon area needed to realise the circuit.     «

More and more embedded applications demand to cope with complex data while still being energy-efficient. Neural networks provide the processing capabilities, but often cannot be utilised because of restricted power goals. Spiking neural networks have been shown to potentially solve this problem due to their hardware friendliness and energy efficiency. One remaining problem is the conversion of input data into event-based spikes in order to be processed. In this study, we examine using resonating...     »