Empowering On-Device Model Adaptation with an Edge AI Inference Accelerator
The paper uses Hailo-8L as a frozen feature extractor so only a small classifier head trains on the device.
The backbone is quantized to INT8 and run on the edge accelerator, while the host CPU fine-tunes a lightweight FP32 head. The authors report up to 15.4x faster wall-clock training than a Raspberry Pi 5 CPU baseline, with lower energy per sample. They also find post-training quantization restoration is important for keeping feature quality in more sensitive architectures. ArXiv · AI/CL/LG's note
The backbone is quantized to INT8 and run on the edge accelerator, while the host CPU fine-tunes a lightweight FP32 head. The authors report up to 15.4x faster wall-clock training than a Raspberry Pi 5 CPU baseline, with lower energy per sample. They also find post-training quantization restoration is important for keeping feature quality in more sensitive architectures. ArXiv · AI/CL/LG's note
score 4