A Rapid Pipeline for Training and Deploying ML Models on WeBe Band

Developing machine learning (ML) models for resource-constrained edge devices like the WeBe Band typically requires significant expertise in both data science and low-level embedded systems. This creates a bottleneck for researchers in fields like healthcare who want to deploy models on wearables. The authors address this by creating an automated pipeline that abstracts away the complexities of firmware development, allowing researchers to focus on model iteration and hardware-aware evaluation.

The proposed framework integrates the Piccolo AI ecosystem with a custom backend that automates the entire deployment lifecycle. After training a model, the system generates the necessary header files, compiles the firmware, and packages it for over-the-air (OTA) delivery. This allows users to flash models directly onto the WeBe Band—a device powered by an ARM Cortex-M4F microcontroller—without manually interacting with the device's underlying firmware or compiler toolchains. The researchers evaluated several model types, including Random Forest, a Pattern Matching Engine (PME), and various neural network architectures, using on-device profiling to measure inference latency and memory usage.

The study demonstrates that while neural networks offer higher representational capacity, they incur a significantly larger memory footprint and higher inference latency compared to classical models like Random Forest and PME. The automated pipeline successfully enabled real-time gesture recognition on the WeBe Band, with all tested models operating within the device's memory and power constraints. The results highlight that classical ML models provide a superior balance of performance and resource efficiency for this specific hardware, while also proving that the automated workflow can successfully bridge the gap between model development and real-world deployment.