Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces QScheduler, an adaptive algorithm for Zeroth-Order optimization that dynamically adjusts the number of gradient samples during on-device training on INT8 NPUs, specifically tested on the STM32N6's Neural-ART NPU. By optimizing the sampling strategy, QScheduler mitigates the trade-off between gradient noise and computational efficiency, achieving performance comparable to well-tuned fixed-sample configurations without the need for extensive hyperparameter tuning. Experimental results on EuroSAT and STL-10 demonstrate that QScheduler effectively enhances on-device learning capabilities while maintaining resource efficiency.
QScheduler achieves optimal gradient sampling dynamically, eliminating the need for costly hyperparameter searches in on-device training.
Zeroth-Order (ZO) optimization enables On-Device Learning (ODL) on NPU-equipped microcontrollers by estimating gradients through forward passes alone, bypassing the need for backpropagation primitives and reducing memory requirements. The number of gradient samples q critically affects training: insufficient samples produce noisy gradients that plateau early, while excessive samples consume more computational resources. However, finding an optimal q typically requires costly hyperparameter searches. This work introduces QScheduler, an adaptive algorithm that adjusts q based on training progress, and provides the first proof-of-concept of INT8 quantized on-device training on the STM32N6's Neural-ART NPU. Experiments on EuroSAT and STL-10 show that QScheduler matches well-tuned fixed-q configurations for both ResNet18 and MobileNetV2, without requiring prior q hyperparameter optimization.