Keeping a microcontroller learning after it ships
Small devices are usually trained once and then frozen. My research on continual learning asks how a microcontroller can keep adapting to new data in the field, without forgetting what it already knows, inside a memory budget of a few hundred kilobytes.
I study this at the CCIS Laboratory, National Formosa University (Taiwan), funded by a Taiwan Ministry of Education scholarship (FTIP 2026), on an STM32F746-class board. Earlier work budgets the memory for sparse weight updates and for a replay of past examples separately. I treat them as one shared pool. How that pool is divided matters a great deal when memory is tight and much less when it is generous, and accuracy keeps improving in a steady, predictable way as the replay grows.
Getting there meant rebuilding the TVM/TinyEngine toolchain, fixing bugs along the way, and finding that memory use measured on the device disagrees with the analytical cost model. On tiny hardware, nothing should be trusted until it is measured.
Pronunciation feedback for deaf and hard-of-hearing learners
A second project, in the Department of Electrical Engineering at National Formosa University with Prof. Hui-Kai Su and an English-studies phonetician, asks a different question: how can a learner who cannot rely on hearing see whether a consonant was pronounced correctly? I am the technical lead on a prototype that finds the consonant in a recording, measures it, and answers with simple colour-coded visual feedback. It is ongoing, and a first demo is planned for December 2026.
Earlier, at TINH VI Lab (HUFLIT), I deployed models on constrained hardware: real-time traffic-sign detection, Android footwear defect inspection, thermal-vision hydrogen-leak monitoring, and edge-AI retrieval for Vietnamese traffic law.
Nguyen Minh Quang
