Supplementary Material On-Device Training Under 256KB Memory

We prepared a video demo showing that we can deploy our framework to a microcontroller (STM32F746, 320KB SRAM, 1MB Flash) to enable on-device learning. The training leads to decent accuracy within the tight memory budget. This is quite affordable for tiny on-device learning applications. We notice that the variance of different runs is quite small in our experiments. Here we provide the results of 3 runs in Table S1 to show the variance.