# QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUs

arXiv: https://arxiv.org/abs/2607.18802 · HTML: https://mlsanity.com/paper/2607.18802/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 10 of its 22 references that are in the map.
- First posted on arXiv 21 Jul 2026.

## Builds on (most cited first, 10 of 10)

- [Adam: A Method for Stochastic Optimization](https://mlsanity.com/paper/1412.6980/) (2014)
- [Training Deep Nets with Sublinear Memory Cost](https://mlsanity.com/paper/1604.06174/) (2016)
- [A Survey of Quantization Methods for Efficient Neural Network Inference](https://mlsanity.com/paper/2103.13630/) (2021)
- [Fine-Tuning Language Models with Just Forward Passes](https://mlsanity.com/paper/2305.17333/) (2023)
- [MCUNet: Tiny Deep Learning on IoT Devices](https://mlsanity.com/paper/2007.10319/) (2020)
- [On-Device Training Under 256KB Memory](https://mlsanity.com/paper/2206.15472/) (2022)
- [MCUNetV2: Memory-Efficient Patch-based Inference for Tiny Deep Learning](https://mlsanity.com/paper/2110.15352/) (2021)
- [TinyOL: TinyML with Online-Learning on Microcontrollers](https://mlsanity.com/paper/2103.08295/) (2021)
- [POET: Training Neural Networks on Tiny Devices with Integrated Rematerialization and Paging](https://mlsanity.com/paper/2207.07697/) (2022)
- [Stepping Forward on the Last Mile](https://mlsanity.com/paper/2411.04036/) (2024)

## Abstract (arXiv)

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.