Beam Search Optimizes Split Learning Latency on ESP32-S3
A new paper presents the first experimental latency benchmark of TinyML-based split learning on ESP32-S3 boards, comparing UDP, TCP, ESP-NOW, and BLE protocols. The study analyzes split point choices across MobileNet-V2 and ResNet50 to minimize end-to-end inference latency. A Beam Search-based algorithm for split point optimization is proposed and compared with Greedy Search and other methods.
Key facts
- First experimental latency benchmark of TinyML-based SL on ESP32-S3 boards
- Compared four wireless protocols: UDP, TCP, ESP-NOW, BLE
- Analyzed split points across MobileNet-V2 and ResNet50
- Proposed Beam Search-based algorithm for split point optimization
- Compared with Greedy Search and other methods
- Published on arXiv with ID 2507.16594
- Split learning addresses deep learning inference on low-power edge/IoT nodes
- Inference latency under realistic low-power wireless protocols was previously unexplored
Entities
Institutions
- arXiv