Spiking Neural Circuitry for Real-Time Decision Making in Autonomous Edge Devices with Energy Constraints

Abstract

Spiking Neural Networks (SNNs) have emerged as a biologically inspired and energy-efficient solution for intelligent decision-making, particularly in low-power edge devices. However, traditional SNNs struggle with learning stability, latency control, and seamless deployment in autonomous platforms due to their asynchronous nature and limited adaptability. These limitations hinder their adoption in real-time applications that demand both computational efficiency and robustness. This study proposes a novel spiking neural circuit designed to support real-time decision-making in energy-constrained edge environments. The architecture integrates temporally encoded input processing, spike-timing dependent plasticity (STDP) for learning, and an adaptive thresholding mechanism to dynamically regulate energy consumption. The model was validated using the Neuromorphic MNIST (N-MNIST) dataset, which simulates real-world event-based sensory inputs. The results demonstrate that the proposed SNN achieves a classification accuracy of 93.1%, with an average inference latency of 17.5 ms and energy consumption as low as 2.3 mJ per inference. Compared to conventional ANN and CNN baselines, this architecture yields a 4× improvement in energy efficiency and a 1.7× reduction in latency, without significant compromise in predictive performance. Performance metrics were further supported through 5-fold cross-validation and simulated Loihi hardware profiling. This research contributes a scalable, low-power, and real-time SNN solution tailored for edge AI deployment. Its implications span across autonomous navigation, biomedical monitoring, and smart surveillance, demonstrating practical feasibility for neuromorphic intelligence in embedded systems.