Memory-Latency-Accuracy Trade-Offs for Continual Learning on a RISC-V Extreme-Edge Node