Summary of On-device Training Under 256kb Memory, by Ji Lin et al.
On-Device Training Under 256KB Memory
by Ji Lin, Ligeng Zhu, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, Song Han
First submitted to arxiv on: 30 Jun 2022
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes an innovative framework for on-device training of convolutional neural networks (CNNs) with minimal memory resources. The key challenge is fine-tuning a pre-trained model using only 256KB of memory, which is prohibitive for Internet of Things (IoT) devices. To overcome this limitation, the authors develop an algorithm-system co-design framework that combines Quantization-Aware Scaling and Sparse Update to optimize neural networks at low bit-precision and limited hardware resources. The proposed Tiny Training Engine prunes the backward computation graph to support sparse updates and offloads runtime auto-differentiation to compile time. This framework enables IoT devices to perform not only inference but also continuous adaptation to new data for on-device lifelong learning, matching the accuracy of PyTorch and TensorFlow while using less than 1/1000th of their memory. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary On-device training lets machines learn from new data without sharing it with others. This is important because it keeps the data private. However, current solutions require a lot of memory, which isn’t available on tiny devices like those used in IoT. The authors come up with a clever solution that uses only 256KB of memory to train AI models. They do this by developing an algorithm and system together that makes it possible to fine-tune pre-trained models without using too much memory. |
Keywords
* Artificial intelligence * Fine tuning * Inference * Precision * Quantization
