Enabling Efficient Integer-Only Few-Shot Learning on Edge Devices

On-device neural network training has emerged as a key enabler for privacy-preserving and personalized data processing at the edge. However, the deployment of Few-Shot Learning (FSL) methods on resource-constrained devices remains underexplored, despite its strong potential for rapid adaptation with minimal data. We present an efficient on-device FSL approach termed Integer-Only Few-shot Adaptation (IOFA), a fully quantized fine-tuning pipeline covering forward, backward, and update phases. Experiments on MiniImageNet show that fully-static INT8 meta-test achieves 45.01% accuracy in 1-shot tasks and 61.36% accuracy in 5-shot settings, with accuracy losses of only 2.34% and 1.75%, respectively, compared to a floating-point implementation. We preserve few-shot accuracy by performing all meta-test operations in integer arithmetic while drastically reducing memory and compute demands, enabling practical deployment on memory-limited embedded systems.

IEEE Embedded Systems Letters