Variable-precision neuromorphic state space model for on-edge activity classification

Abstract

Neuromorphic computing is rising as a promising paradigm for efficient AI, leveraging event-driven computation to achieve low-power and high-performance computing. Due to the real-time processing required by edge devices with minimal power consumption, optimizing neuromorphic models for on-edge applications can be crucial to address the issue of power efficiency and resource-constraint devices. This work explores the definition of a neuromorphic state space model and its deployment on non-dedicated hardware. Structured sparsity and quantization techniques are leveraged to enhance the model’s efficiency. By compressing synaptic operations and memory footprint, we demonstrate how neuromorphic models can be adapted for on-edge deployment, ensuring low-latency and memory efficient inference. This study highlights the potential of neuromorphic models as a scalable solution for real-world embedded systems with limited resources.