Item

Abstract

Realistic simulations of spiking neural networks are computationally very costly since each modeled neuron usually requires updates of multiple ODEs [1] in every computational step. A number of proposed approaches to achieving reasonable simulation times for spiking networks rely on various optimization techniques. A radical approach to improving computational efficiency is to employ cellular automata (CAs) that utilize the power of large-scale parallelization and optimized matrix operations to achieve real-time simulations of extensive systems [2]. Cellular Automata are a configuration of cells, usually arranged in a grid, whose state changes according to a local update rule applied globally. Due to the locality of the update rule, multiple cells can be updated in parallel, allowing for distributed computing. Describing a spiking network as a CA enables leveraging parallelization for fast computations [3].



Here, we develop a CA model that aims to simulate the activity of a balanced spiking neural network. Using multiple CA channels and optimized multi-layered connectivity kernels, we include various features of biological neural networks such as axonal delays, energetic constraints [4], and homeostatic plasticity (threshold adaptation). We demonstrate that our network exhibits asynchronous-irregular dynamics and develops realistic distributions of firing rates and CVs of the Inter-spike intervals. Additionally, we demonstrate that the threshold adaptation allows for the imprinting of short-term memories in the activity of the network, replicating the performance of simulated spiking networks in similar tasks [5]. Finally, we use our neural-CA as a reservoir computing model [6]. We find that its dynamics can be used to accurately predict the trajectory of a chaotic time series (tested and illustrated for the Lorenz attractor) using previous time steps as external input. Our findings indicate the usefulness of CAs for very efficient and biologically realistic spiking network simulations that accurately capture many of the dynamical and computational properties of spiking neural networks.