A new technique enables on-device training of machine-learning models on edge devices like microcontrollers, which have very limited memory. This could allow edge devices to continually learn from new data, eliminating data privacy issues, while enabling user customization.