FPT-spike: a flexible precise-time-dependent single-spike neuromorphic computing architecture
- PDF / 2,855,450 Bytes
- 18 Pages / 595.276 x 790.866 pts Page_size
- 56 Downloads / 188 Views
REGULAR PAPER
FPT‑spike: a flexible precise‑time‑dependent single‑spike neuromorphic computing architecture Tao Liu1 · Gang Quan2 · Wujie Wen1 Received: 11 March 2020 / Accepted: 14 May 2020 © China Computer Federation (CCF) 2020
Abstract Modern Artificial Neural networks (ANNs) like Convolutional Neural Network (CNN), have found broad applications in real-world cognitive tasks. One challenging faced by these models is their tremendous memory and computing resource requirement. This also greatly hinders their adoptions from resource-constraint platforms, such as drone, mobile phone and IoT devices. Recently the brain-inspired spiking neural network (SNN) has been demonstrated as a promising solution for delivering more impressive computing and power efficiency. For SNNs, a large body of prior work were conducted on the spiking system design with a focus on using the spike firing rate (or rate-coded) for fulfilling the practical cognitive tasks. Such rate-based designs can underestimate the energy efficiency, throughput and system flexibility of SNNs. On the other hand, the potentials of time-based SNN are not fully unleashed in real applications due to lack of efficient coding and practical learning schemes in temporal domain. In this work, we make an early attempt to fill this gap: that said, a flexible precisetime-dependent single-spike neuromorphic computing architecture, namely “FPT-Spike”, is developed. “FPT-spike” relies on three hardware-favorable components: precise ultra-sparse spike temporal encoding, efficient supervised temporal learning and fast asymmetric decoding, to realize flexible spatial–temporal information trade-off for neural network size reduction without scarifying data processing capability. Extensive experimental results show that “FPT-Spike” outperforms rate-based SNN and ANN significantly in three aspects: network size, processing speed and power consumption, demonstrating great potentials for its deployment in edge devices. Keywords Neuromorphic computing · Spiking neural network · Time coding
1 Introduction As one of the most fascinating developments of Artificial Intelligence (AI), deep learning enabled neural network system, i.e. deep neural network (DNN) or convolutional This work is supported in part by NSF under project CNS1423137. * Tao Liu [email protected] Gang Quan [email protected] Wujie Wen [email protected] 1
Department of Electrical and Computer Engineering, Lehigh University, Bethlehem, USA
Department of Electrical and Computer Engineering, Florida International University, Miami, USA
2
neural network (CNN), has found broad applications in realistic cognitive tasks such as speech recognition, image processing, machine translation and object detection etc. (LeCun et al. 2015; Szegedy 2016). However, performing high-accurate testings for complex DNNs or CNNs requires massive amounts of computation and memory resources, indicating limited energy efficiency. For instance, the recognition implementation of state-of-the-art CNN–AlexNet (Krizhevsky et al. 2012) i
Data Loading...
