qLIF: Mitigating the memory and computation overhead to implement spiking convolutional neural networks

qLIF: Mitigating the memory and computation overhead to implement spiking convolutional neural networks

연구 분야: Artificial Intelligence

학회: Neural Computing and Applications

Spiking neural networks (SNNs) have many advantages in achieving lower power neuromorphic computation. In this paper, we propose new methods that can further mitigate the memory spaces and computation resources that are needed for spiking convolutional neural networks (SCNNs) implementation. Traditional leaky-integrate-and-fire (LIF) mechanism, which is widely adopted in SNN, demands large memory spaces to store the membrane potentiation values for integration, through experiments, we find out that for many layers, the potentiation values can be quickly leaked (qLIF) and no extra memory spaces are needed to store the values from past events. With the adoption of qLIF mechanism, convolution between the input matrix and the kernel matrix can be simplified into a look-up-table (LUT) operation (called ConvLUT in this paper), without the need of computation resources to perform the dot-product. Through experiments on MNIST, CIFAR-10, CIFAR-100 and DVS128-Gesture datasets with SCNNs, we have verified that both qLIF mechanism and ConvLUT method can greatly reduce the memory spaces as well as computation resources in SCNN, the computation latency can also be improved.