Neuromorphic Computing for Energy-Efficient Artificial Intelligence

Abstract

Neuromorphic computing draws inspiration from the structure and dynamics of biological neural systems to deliver radically more energy-efficient artificial intelligence than conventional von-Neumann hardware. By representing information as sparse, asynchronous spikes and co-locating memory with computation, neuromorphic systems achieve orders-of-magnitude reductions in power for specific workloads such as event-based vision, sensor fusion, and always-on inference. This paper reviews the principles of spiking neural networks (SNNs), surveys major hardware platforms TrueNorth, Loihi 1 and 2, SpiNNaker 1 and 2, BrainScaleS, and Akida and analyses learning rules ranging from biologically plausible spike-timing-dependent plasticity to surrogate-gradient back-propagation. Applications, software ecosystems, evaluation benchmarks, and persistent challenges are discussed.