Continual Learning on TinyML Hardware: Benchmarking Study

Evaluating Streaming Continual Learning on TinyML Hardware Under Concept Drift

MSc Thesis Proposal

Alba Huti

JKU – Streaming AI Initiative

PROBLEM

STATEMENT

Static TinyML Models

Most deployed TinyML models are trained offline and cannot adapt after deployment. When input distributions drift, accuracy collapses.

Concept Drift in Industrial IoT

Tiny embedded devices (MCUs) sense changing environments (e.g. worker activity, machine wear), but static models fail to handle these shifts.

Costly Cloud Retraining

Current workaround is expensive cloud-based retraining, which contradicts goals of sustainability, privacy, and edge autonomy.

No Systematic Benchmark Exists

No existing work has comprehensively evaluated multiple lightweight online-learning methods on real MCUs under drift simultaneously measuring accuracy, latency, and power.

Research Objective

Design and benchmark lightweight online continual learning methods that run entirely on microcontroller-class hardware — enabling on-device adaptation to concept drift without cloud retraining.

01

Algorithm Implementation

SGD Perceptron, Experience Replay & EWC-lite deployed on Arm Cortex-M4 MCU

02

Real-World Evaluation

UCI HAR sensor data streamed sequentially with controlled distribution shifts

03

Reproducible Framework

Trade-off curves across accuracy · memory · latency · energy

EXPECTED OUTCOME

70% accuracy at 10KB RAM

85% recovery

Research Questions

Which lightweight continual-learning methods are most effective at maintaining predictive performance under concept drift on microcontroller-class TinyML hardware?

What trade-offs arise between predictive performance, memory usage, latency, and energy consumption when implementing streaming learning on resource-constrained MCUs?

How much additional computational and memory overhead is required for continual-learning methods to recover from concept drift compared with static TinyML models?

04

Methodology

Slide 05