On-Device Transfer Learning based on Mixed Precision Partitioning
Keywords: Machine Learning, Transfer Learning, On-device training
TL;DR: This paper aims to improve the efficiency and performance of transfer learning in embedded systems.
Abstract: The application of machine learning is becoming more widespread, with a growing number of use cases. The development of centralized data training and the exponential growth of data generation raise significant privacy and security concerns. On-device training offers a solution by enhancing privacy and reducing the need for communication between the cloud and the device. Furthermore, on-device transfer learning (TL) can leverage the knowledge gained from pre-trained models, hence, accelerating the training process. However, backpropagation, especially in embedded systems, requires more memory than running inference, which becomes a challenge for devices with limited resources. This paper aims to improve the efficiency and performance of on-device TL. We propose an open source mixed-precision partitioning framework that identifies optimal partitioning layers for retraining, combining quantized and bfloat16 layers to enhance performance and energy efficiency. Our approach is validated through experiments on ResNet-18 and SqueezeNetV1.1 models using Flowers-102, STL-10, and OxfordIIITPet datasets. The partitioned mixed-precision model is able to transfer the knowledge from the pre-trained model to new datasets without losing accuracy compared to the baseline bfloat16 model. These results illustrate the potential for resource-constrained devices to perform TL locally.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10529
Loading