What changes after deployment? A survey on On-device Learning in TinyML
A comprehensive survey of on-device learning systems reveals how TinyML deployments must adapt to real-world distribution shifts after launch. By analyzing 70 existing solutions through the lens of different change regimes, researchers expose a critical gap between controlled benchmarks and actual field conditions. This work matters for practitioners building edge AI: it clarifies which hardware, model architectures, and learning strategies suit specific drift patterns, directly informing how to architect systems that remain effective as user data diverges from training distributions.
Modelwire context
ExplainerThe survey's real contribution isn't that drift exists after deployment (practitioners know this), but that it systematically maps which learning strategies, hardware constraints, and model sizes actually handle specific drift regimes in the field. Most TinyML work optimizes for inference latency on static models; this work exposes how few solutions address the harder problem of continuous adaptation on microcontrollers with severe memory and power budgets.
This is largely disconnected from recent activity in the broader ML deployment space, which has focused on larger models and post-hoc alignment. The relevant lineage is narrower: embedded systems and edge ML have long struggled with the gap between controlled benchmarks and real sensor data, but most prior work either ignored drift entirely or assumed cloud retraining was available. This survey belongs to the growing body of work asking what happens when you can't phone home, which matters most in robotics, IoT, and industrial monitoring where connectivity is unreliable or forbidden.
If practitioners adopt the survey's taxonomy to design new TinyML systems and report field performance metrics (not just benchmark scores) within the next 18 months, that signals the work moved beyond academic categorization into actual design guidance. Conversely, if the 70 solutions reviewed remain largely unchanged in their deployment strategies, the survey is documentation of the status quo rather than a catalyst for change.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsTinyML · On-device Learning · Microcontroller-class devices
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.