Digital neuromorphic real-time platform

• Published in: Neurocomputing (Amsterdam) Vol. 371; pp. 91 - 99
• Main Authors: Perez-Peña, Fernando, Cifredo-Chacon, M. Angeles, Quiros-Olozabal, Angel
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 02.01.2020
• Subjects: Brian simulator; FPGA; Leaky-integrate and fire; Neuromorphic engineering; Spiking neural network; Leaky-integrate and fire; Neuromorphic engineering; Spiking neural network; FPGA; Brian simulator
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2019.09.004

Hardware implementations of spiking neural networks in portable devices can improve many applications of robotics, neurorobotics or prosthetic fields in terms of power consumption, high-speed processing and learning mechanisms. Analog and digital platforms have been previously proposed to run these networks. Analog designs are closer to biology since they implement the original mathematical model. However, digital platforms are, to some extent, abstractions of this model so far. In this paper, a full digital platform to design, implement and run real-time analog-like spiking neural networks is presented. Specifically, we present the design and implementation of digital circuits to run real-time biologically plausible spiking neural networks on a Field Programmable Gate Array (FPGA). The circuit designed for the neuron implements the Leaky Integrate and Fire (LIF) model. The synapsis implemented is a bi-exponential current-based one. The synaptic circuit design consists of one static memory with the baseline current and a dynamic memory which stores the updated contribution over time of each pre-synaptic connection. All the parameters of both the neuron and the synapse are configurable. The results of the circuits are validated by running the same experiments on the Brian simulator. The circuits, which are totally original and independent of the technology, use only 136 slice registers of hardware resources. Thus, these designs allow the scale of the network. These circuits aim to be the basis of the spiking neural networks on digital devices. This platform allows the user to first simulate their network within the Brian simulator and then, confidently, move to the hardware platform replicating the same performance or even replace their analog platform with the digital one.