Progress and Challenges in Large Scale Spiking Neural Networks for AI and Neuroscience

Abstract

The growing demand for energy-efficient and biologically plausible artificial intelligence has driven significant interest in neuromorphic computing and event-driven neural processing. Among neuromorphic approaches, spiking neural networks (SNNs) have emerged as a compelling alternative to traditional deep learning models, offering advantages in temporal information processing, low-power computation, and real-time adaptability. However, despite their potential, scaling these networks to large, biologically realistic architectures remains a fundamental challenge due to constraints in training methodologies, hardware limitations, and computational complexity. This paper provides a comprehensive survey of large-scale SNNs, covering key aspects such as computational models, network architectures, and advancements in supervised, unsupervised, and reinforcement learning methods. We discuss the latest progress in neuromorphic hardware, including digital, analog, and hybrid implementations, which facilitate efficient execution of large-scale SNNs. Furthermore, we explore real-world applications, from robotics and brain-computer interfaces to edge computing and event-based vision, highlighting the advantages and practical constraints of SNN-based solutions. In addition to surveying existing research, this paper identifies key challenges in scalability, training efficiency, and hardware integration, offering insights into potential future directions. By addressing these limitations and leveraging interdisciplinary innovations, large-scale SNNs hold the promise of bridging the gap between artificial intelligence and brain-like computation, paving the way for next-generation intelligent systems.