MetaCLBench: Benchmarking the On-Device Adaptation Cost of Meta-Continual Learning

Meta-Continual Learning (Meta-CL) enables models to learn new classes from a few labelled samples, which makes it attractive for IoT applications where labelling is costly. However, existing evaluations report accuracy and largely ignore whether adaptation is feasible on resource-constrained hardware. We present MetaCLBench, a benchmark that evaluates Meta-CL methods for accuracy together with the deployment-critical costs of learning (peak memory, end-to-end latency and energy) on real edge devices with 512 MB to 4 GB of RAM. We evaluate six Meta-CL methods across three architectures (CNN, YAMNet, ViT) and five image and audio datasets. Depending on the dataset, up to three of the evaluated methods run out of memory on a 1 GB device, which sharply narrows the set of deployable options. LifeLearner reaches near-oracle accuracy while using 2.54-7.43x less energy than the oracle baseline, and larger pretrained architectures do not consistently improve the Meta-CL accuracy--efficiency trade-off. We distil these findings into deployment guidelines and release the framework at https://anonymous.4open.science/r/MetaCLBench-874E/.