# Deep Learning on Mobile Devices Through Neural Processing Units and Edge Computing

arXiv: https://arxiv.org/abs/2112.02439 · HTML: https://mlsanity.com/paper/2112.02439/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 4 Dec 2021.

## Built on by (most cited first, 5 of 5)

- [Edge Video Analytics: A Survey on Applications, Systems and Enabling Techniques](https://mlsanity.com/paper/2211.15751/) (2022)
- [EdgeWisePersona: A Dataset for On-Device User Profiling from Natural Language Interactions](https://mlsanity.com/paper/2505.11417/) (2025)
- [Monocular Obstacle Avoidance Based on Inverse PPO for Fixed-wing UAVs](https://mlsanity.com/paper/2411.18009/) (2024)
- [Legilimens: Performant Video Analytics on the System-on-Chip Edge](https://mlsanity.com/paper/2504.21136/) (2025)
- [FM-CAC: Carbon-Aware Control for Battery-Buffered Edge AI via Time-Series Foundation Models](https://mlsanity.com/paper/2604.16448/) (2026)

## Abstract (arXiv)

Deep Neural Network (DNN) is becoming adopted for video analytics on mobile devices. To reduce the delay of running DNNs, many mobile devices are equipped with Neural Processing Units (NPU). However, due to the resource limitations of NPU, these DNNs have to be compressed to increase the processing speed at the cost of accuracy. To address the low accuracy problem, we propose a Confidence Based Offloading (CBO) framework for deep learning video analytics. The major challenge is to determine when to return the NPU classification result based on the confidence level of running the DNN, and when to offload the video frames to the server for further processing to increase the accuracy. We first identify the problem of using existing confidence scores to make offloading decisions, and propose confidence score calibration techniques to improve the performance. Then, we formulate the CBO problem where the goal is to maximize accuracy under some time constraint, and propose an adaptive solution that determines which frames to offload at what resolution based on the confidence score and the network condition. Through real implementations and extensive evaluations, we demonstrate that the proposed solution can significantly outperform other approaches.