FPGA implementation of hippocampal spiking network and its real-time simulation on dynamical neuromodulation of oscillations

Highlights

• We address the challenges of engineering the digital hippocampal neural network with oscillation dynamics.

• A new design is proposed to implement the endogenous surroundings and neural heterogeneity within the neural system.

• A novel scalable network-on-chip architecture is proposed for the establishment of randomly connected SNNs.

• The importance of network dynamics is highlighted in the design to be more bio-realistic.

Abstract

Neural information is represented and transmitted among single neurons by a series of all-or-none neural codes with certain oscillation dynamics. Real-time implementation of the hippocampal spiking network is a promising avenue to investigate the complexity underlying spatiotemporal information encoding and the emergent coherence that arises with the properly coupling of large number of neurons. This paper presents a real-time scalable hardware platform for implementing hippocampal spiking neural network (HSNN) with 10 K neurons, which introduces a novel network-on-chip architecture for the randomly connected spiking neural networks (SNNs). The effects of endogenous surroundings and neural heterogeneity are taken into consideration in the hardware design, which replicates more relevant biological dynamics in comparison with the state-of-the-art studies. Based on the hardware synthesis and theoretical analysis, it is demonstrated that the proposed implementation is able to mimic hippocampal oscillation modulation dynamics under external stimuli, which is vital for the reasonable design of noninvasive electrotherapeutic strategies. The proposed implementation is meaningful for both the efficient hardware implementation of the randomly connected SNNs and the dynamic investigation of the HSNNs.

Introduction

Hippocampal oscillations share some important features throughout the entire central nervous system of higher animals, which entrain active neurons to fire action potentials at strictly defined phases of the oscillation cycle. Network oscillations of the hippocampal formation have been one of the main focuses of neuroscience research, given the role of the hippocampus-temporal lobe in cognitive functions [1], [2], [3]. Neurological and psychiatric disorders are associated with impaired neuronal network oscillations, which are considered a critical pathophysiology of clinical symptoms [4], [5]. Electrical neuromodulation of hippocampal oscillations is an approach in the field of biomedical engineering that is promise for the alleviation of symptoms related to neurological disorders such as Parkinson's disease and epilepsy [6]. Although most of the data on hippocampal oscillatory activity has been recorded from electrophysiological recordings of rodents, recent findings also indicate a similar oscillatory activity in the hippocampal formation of the human brain [7], [8]. To foster a further understanding of dynamical oscillation in large-scale hippocampal spiking networks closely related to neurological disorders, a high-performance computational platform that can effectively reproduce the complicated dynamics and the functions of the hippocampal spiking network will be essential.

Some early findings suggested that the hippocampus is primarily involved in spatial processing [9] and the hippocampal rhythm of oscillation supports general neural computational processes in various species that extend far beyond the neural representation of space [10], [11], [12]. Place cells and grid cells also show a form of temporal or phase coding relative to the ongoing theta rhythm of the local field potential (LFP) in addition to their spatially modulated firing rates [13], [14], [15]. Therefore, the neuronal processes and mechanisms underlying neuronal network oscillations have been intensively studied using the computational methods based on the hippocampal spiking neural network (HSNN) [16], [17], [18], [19], [20], [21], [22].

Although even the simplest behaviors in mammals involve many millions of neurons, the previous CPU-based investigations of the oscillation dynamics of the HSNNs using commercial software usually contained small number of cells for the sake of computational time, yet cannot meet the real-time computational requirement that is the prerequisite of neuroprothetic technology. Conventional CPU-based solutions have achieved a significant enhancement on computing power, but there still exist several challenges, such as the difficulty to further improve clock rates and an incongruity between memory bandwidth and ever-increasing CPU speed [23]. VLSI designs are limited by its high costs and long development time limit its applications [24], [25]. Current GPU programming paradigms use a kernel-launch method in which a large number of calculations is offloaded onto the GPU with a batch of data and cannot continuously run. In addition, the power consumption of GPUs hinders the applications [26], [27], [28]. Field-programmable gate array (FPGA) is featured with the reconfigurable structure, parallel calculation and distributed form. Currently, numerous FPGA-based designs of neural networks have been presented for different applications with different structures [29], [30], [31], [32], [33]. Four factors are ignored in the previous implementations of the biologically plausible neural networks. Firstly, few FPGA-based neural network implementations have shown the dynamical oscillations, although it is vital for the information coding and processing in the cognitive system [34]. Secondly, the neural heterogeneity in these studies is always ignored, which reduces the accuracy of the biological dynamics on the network level [35]. On the other hand, the effects of endogenous surroundings around the network are always neglected, which also modulates the network oscillations [36]. Besides, to the best of our knowledge, the challenges of efficient implementation of large-scale SNNs with randomly connected coupling synapses have been rarely addressed. These factors will limit the accurate reproduction of the network dynamics and high-performance computation of the neural systems.

Previous studies have presented the FPGA-based designs for the functional neural network in human brain such as the basal ganglia network and the thalamocortical network [33], [37]. These studies have reproduced the network behaviors of the basal ganglia-thalamus circuit which is closely related to movement disorders. In this study we have focused on the investigations of the real-time implementation of the HSNN with intrinsic oscillations and its neuromodulation of oscillations, and made a further exploration of the system performance. This work presents a critical technique to investigate the dynamical modulation of oscillations in the large-scale hardware-based HSNN, especially under the stimuli of both the endogenous field and the external applied electrical field. In particular, four efforts should be highlighted in this paper. Firstly, a scalable FPGA-based architecture for prototyping the HSNN with biologically realistic neuron and synapse models is implemented, which can regenerate the biological dynamics of the HSNN in real time under both the normal and the oscillation modulation conditions. Secondly, the proposed platform can reveal the hippocampal oscillations under the external electric field stimulus, which is beneficial for the reasonable design of noninvasive electrotherapeutic strategies in a new perspective. Thirdly, the proposed system includes a versatile input/output module to interact with other systems or living cells, which is helpful for the exploration of the closed-loop neural dynamics control. Fourthly, we have compensated the hardware-induced distortion for a high accuracy of dynamics, and given a scalable topology that can be used for other randomly connected spiking neural networks (SNNs) with complicated spiking activities.

The rest of the paper is organized as follows. In Section 2, the computational model of the HSNN is introduced. Hardware design and architecture optimization are explained in Section 3, and the analysis for the dynamics of HSNN system and its network architecture is presented in Section 4. Section 5 illustrates the implementation results with the precision analysis. A discussion will be presented in Section 6, and this paper concludes with a summary of the hardware implementation in Section 7.