Authors: Kolhe Om Dinkar, Prof. R. V. Varhade, Prof. M. C. Kolse, Prof. S. K. Said
Abstract: Edge AI requires intelligent processing on resource-constrained devices such as wearables, drones, smart sensors and Internet of Things nodes. Conventional deep learning on von Neumann architectures consumes significant power because data must move continuously between memory and processing units, limiting its use in battery-powered applications. Neuromorphic computing addresses this challenge by emulating the structure of the biological brain through event-driven, massively parallel processing with memory co-located with computation. At its core are Spiking Neural Networks (SNNs), which communicate through sparse, discrete spikes over time so that energy is consumed mainly when meaningful events occur. This paper reviews the principles of neuromorphic computing and SNNs, covering neuron models such as the leaky integrate-and-fire neuron, spike encoding schemes, and training methods including ANN-to-SNN conversion, surrogate-gradient learning and biologically inspired local rules. Neuromorphic platforms such as IBM TrueNorth, the SpiNNaker system and Intel Loihi are compared, and edge applications in event-based vision, gesture recognition, keyword spotting and biomedical signal processing are examined. A layered architecture for neuromorphic edge AI is proposed, together with research gaps, objectives and a discussion of challenges such as training complexity, toolchain maturity and hardware standardization.
