Evolutionary FPGA-Based Spiking Neural Networks for Continual Learning

Spiking Neural Networks (SNNs) constitute a representative example of neuromorphic computing in which event-driven computation is mapped to neuron spikes reducing power consumption. A challenge that limits the general adoption of SNNs is the need for mature training algorithms compared with other artificial neural networks, such as multi-layer perceptrons or convolutional neural networks. This paper explores the use of evolutionary algorithms as a black-box solution for training SNNs. The selected SNN model relies on the Izhikevich neuron model implemented in hardware. Differently from state-of-the-art, the approach followed in this paper integrates within the same System-on-a-chip (SoC) both the training algorithm and the SNN fabric, enabling continuous network adaptation in-field and, thus, eliminating the barrier between offline (training) and online (inference). A novel encoding approach for the inputs based on receptive fields is also provided to improve network accuracy. Experimental results demonstrate that these techniques perform similarly to other algorithms in the literature without dynamic adaptability for classification and control problems.