Neuromorphic Computing for Ultra-Low-Power Intelligent Systems: Architectures, Learning Paradigms, Applications, and Emerging Challenges

Abstract

The rapid expansion of artificial intelligence (AI) into edge devices, autonomous systems, intelligent sensors, robotics, healthcare platforms, and pervasive Internet-of-Things environments has exposed fundamental limitations in conventional computing architectures. Although contemporary deep neural networks achieve remarkable performance, their dependence on high-volume memory transfers, dense numerical operations, and centralized acceleration creates substantial energy, latency, and bandwidth costs. Neuromorphic computing offers an alternative paradigm in which computation is organized around sparse, asynchronous, event-driven information processing inspired by biological nervous systems. Spiking neural networks (SNNs), specialized memory and synaptic structures, event-driven communication, and tightly coupled sensing and computation constitute the principal elements of this emerging paradigm. This review critically examines neuromorphic computing as a foundation for ultra-low-power intelligent systems, with particular emphasis on architectures, learning paradigms, algorithms, applications, implementation technologies, and unresolved research challenges. The evolution from early neuromorphic circuits to contemporary large-scale platforms such as Loihi, Loihi 2, SpiNNaker 2, BrainScaleS-2, and event-driven sensing-computing systems is analyzed. SNN learning mechanisms, including spike-timing-dependent plasticity, surrogate-gradient optimization, ANN-to-SNN conversion, temporal coding, local learning, and hybrid learning, are compared from algorithmic and hardware perspectives. Emerging technologies involving memristive devices, spintronic elements, two-dimensional materials, analog computation, and in-memory processing are also examined. Particular attention is given to the relationship among spike sparsity, memory locality, communication overhead, accuracy, latency, and energy consumption. Applications in event-based vision, robotics, autonomous navigation, speech recognition, healthcare, industrial monitoring, smart agriculture, edge intelligence, and brain-machine interfaces are reviewed. The analysis identifies training complexity, benchmark inconsistency, hardware heterogeneity, limited software interoperability, device variability, scalability, security, interpretability, and sustainability as persistent barriers. The review concludes that practical progress will depend less on reproducing the brain literally and more on systematic algorithm–architecture–device co-design, standardized evaluation methodologies, adaptive learning, interoperable software abstractions, and application-specific neuromorphic system design.