# Continual Learning on TinyML Hardware: Benchmarking Study
> Explore a benchmark study on lightweight online learning for microcontrollers, focusing on concept drift adaptation in Industrial IoT and edge devices.

Tags: tinyml, continual-learning, microcontrollers, edge-ai, iot, machine-learning, concept-drift, benchmarking
## Evaluating Streaming Continual Learning on TinyML Hardware Under Concept Drift
- **Thesis Proposal by:** Alba Huti
- **Affiliation:** JKU – Streaming AI Initiative

## Problem Statement
- **Static TinyML:** Models are typically trained offline and fail when input distributions shift (concept drift).
- **Industrial IoT:** Embedded MCUs in changing environments (machine wear, worker activity) experience accuracy collapse with static models.
- **Sustainability:** Cloud retraining is costly and contradicts privacy and edge autonomy goals.
- **Research Gap:** Lack of systematic benchmarks measuring accuracy, latency, and power for online learning on real MCUs.

## Research Objective
- Design and benchmark lightweight online continual learning (CL) methods for microcontroller-class hardware.
- **Implemented Algorithms:** SGD Perceptron, Experience Replay, and EWC-lite on Arm Cortex-M4.
- **Expected Outcome:** ~70% accuracy at 10KB RAM for Perceptron; 85% recovery for Experience Replay.

## Research Questions
- Which CL methods maintain predictive performance best under drift on TinyML hardware?
- What are the trade-offs between performance, memory, latency, and energy?
- What is the computational overhead compared to static TinyML models?

## Methodology
- **Hardware:** Arm Cortex-M4 (Arduino Nano 33 BLE Sense), 64 MHz, 256 KB SRAM, 1 MB Flash.
- **Data:** UCI HAR Dataset (Accelerometer, 50 Hz, 128-sample windows) streamed sequentially.
- **Drift Scenarios:** Testing Sudden, Gradual, Incremental, and Recurring drift types.
- **Metrics:** Prequential accuracy, drift recovery speed, latency, RAM/Flash usage, and energy consumption.
---
This presentation was created with [Bobr AI](https://bobr.ai) — an AI presentation generator.