Основні поняття
MicroT introduces a low-energy, multi-task adaptive model framework for resource-constrained MCUs, improving model performance and reducing energy consumption by implementing stage-decision between part and full models.
Анотація
MicroT proposes a novel approach to address the challenges of deploying DNNs on resource-constrained devices. By dividing the model into a feature extractor and classifier, utilizing self-supervised knowledge distillation, and implementing joint training with stage-decision, MicroT achieves significant improvements in accuracy and energy efficiency. The experiments demonstrate the effectiveness of MicroT in enhancing model performance for multiple local tasks on MCUs while reducing energy consumption.
Key points:
Proposal of MicroT for low-energy, multi-task adaptive models on MCUs.
Division of model into feature extractor and classifier.
Utilization of self-supervised knowledge distillation and joint training with stage-decision.
Significant improvements in accuracy and energy efficiency demonstrated through experiments.
Статистика
MicroT can improve accuracy by up to 9.87% compared to unoptimized feature extractors.
Energy consumption savings of up to about 29.13% on MCUs with MicroT compared to standard full-model inference.