Neuromorphic Computing Architectures for Future Intelligent Systems
DOI:
https://doi.org/10.5281/zenodo.19152670Keywords:
neuromorphic computing, spiking neural networks, brain-inspired AI, Intel Loihi, surrogate gradient, spike coding, energy-efficient AIAbstract
Neuromorphic computing — inspired by the structural and functional organisation of biological neural circuits — offers a radical departure from von
Neumann architectures by co-locating memory and computation, processing information through event-driven spike-based communication, and
achieving orders-of-magnitude improvements in energy efficiency for temporal pattern recognition tasks. Despite significant hardware advances,
systematic frameworks for programming, training, and deploying spiking neural networks (SNNs) across heterogeneous neuromorphic platforms
remain immature. This study presents NeuroFrame, a unified neuromorphic computing framework integrating surrogate gradient training for deep
SNNs, a hardware abstraction layer (HAL) enabling cross-platform deployment, and a temporal coding optimiser for spike-rate efficiency
maximisation. NeuroFrame was evaluated on Intel Loihi 2, IBM TrueNorth, and a custom FPGA neuromorphic accelerator across four intelligent
system benchmarks: N-MNIST digit recognition, DVS-Gesture hand gesture recognition, SHD spoken digit classification, and autonomous drone
obstacle avoidance in simulated cluttered environments. NeuroFrame achieved accuracy improvements of 4.8–11.3% over existing SNN training
methods, energy reductions of 94.7% versus GPU-based ANN inference, and a 3.7× throughput improvement over single-platform SNN deployment.
These results advance neuromorphic computing toward practical deployment in future energy-efficient intelligent systems.
