Advances in Large-Scale Spiking Neural Networks: Learning, Simulation, and Deployment

Spiking Neural Networks (SNNs) represent a biologically inspired class of artificial neural models that leverage discrete spike events to encode and process information with high temporal precision and energy efficiency. Unlike conventional artificial neural networks that rely on continuous-valued activations and synchronous computation, SNNs operate asynchronously through sparse spike trains, closely mimicking the event-driven dynamics of biological neurons. This unique computational paradigm has sparked significant interest for developing large-scale neural models capable of real-time processing under strict power and latency constraints. This survey provides a comprehensive and mathematically rigorous overview of the state-of-the-art in large-scale SNN research, encompassing theoretical foundations, learning algorithms, hardware architectures, software infrastructures, benchmarking methodologies, and application domains. We begin by formalizing neuron and synapse models, highlighting challenges related to non-differentiability of spike functions and temporal credit assignment, and reviewing gradient-based and biologically plausible learning frameworks such as surrogate gradients and spike-timing dependent plasticity. Subsequently, we analyze specialized neuromorphic hardware platforms—including Intel Loihi, IBM TrueNorth, and analog systems—and scalable software simulators, emphasizing their architectural features and computational trade-offs. A critical examination of benchmarking datasets and multi-dimensional evaluation metrics reveals the complexity of assessing large-scale SNNs in terms of accuracy, latency, energy consumption, and biological plausibility. Furthermore, we discuss key applications ranging from robotics and sensory processing to brain-machine interfaces and edge AI, illustrating the advantages and current limitations of large-scale SNN deployment. Finally, we identify open challenges and future research directions, underscoring the importance of hardware-software co-design, standardized benchmarks, and hybrid learning approaches to unlock the full potential of large-scale spiking networks. This survey aims to serve as a foundational reference for researchers and practitioners seeking to advance the design, implementation, and application of scalable, efficient, and biologically grounded neural computation.