Benchmarking TinyML Continual Learning under Concept Drift
Explore a research framework for evaluating streaming AI on ARM Cortex-M4 hardware, focusing on accuracy, RAM, and energy trade-offs for on-device learning.
Evaluating Streaming Continual Learning on True TinyML Hardware under Concept Drift
MSc Thesis Proposal
Alba Huti
JKU — Streaming AI Initiative
Proposed Analysis Tools
Calendar of Activities (Gantt Chart)
A structured timeline breaking the thesis into phases: Literature Review, Implementation, Experimentation, Analysis & Writing. Tracks milestones and ensures timely delivery across the 6-month research period.
Expected Contribution
Benchmark Framework
Open-source evaluation suite for TinyML continual learning on MCUs
Empirical Trade-off Data
Accuracy vs. RAM, latency, energy for each algorithm
Practical Design Guidelines
Concrete recommendations for resource-constrained deployments
Streaming AI Alignment
Supports on-device learning to reduce cloud dependency and CO₂
Thank You
Research at a Glance
🔬
2–3 Continual Learning Algorithms
Online SGD, Replay Buffer, EWC-lite
⚙️
ARM Cortex-M4 MCU
256KB RAM · 64MHz · Real TinyML Hardware
📊
4 Metrics Evaluated
Accuracy · Latency · Memory · Energy
Advancing autonomous, resource-efficient edge intelligence — one sample at a time.
Alba Huti
JKU Streaming AI Initiative
2026
- tinyml
- continual-learning
- concept-drift
- edge-ai
- arm-cortex-m4
- machine-learning-benchmark
- embedded-systems