Toward storage-aware learning with compressed data an empirical exploratory study on JPEG

On-device machine learning is fundamentally constrained by limited storage, especially in continuous data collection scenarios where sensor or vision streams accumulate rapidly. This paper empirically investigates storage-aware learning, characterizing the trade-off between data quantity and data quality under lossy compression. Using the CIFAR-10 dataset as a controlled benchmark, we systematically vary both the amount and the fidelity of training data to understand their joint impact on model performance. Our results reveal that (1) neither maximizing quantity nor quality alone yields optimal accuracy, emphasizing that the optimal trade-off between them depends nonlinearly on the available storage budget, and (2) data samples exhibit differential sensitivity to compression, motivating a sample-wise adaptive compression policy. These findings challenge uniform data-retention strategies such as naive data dropping or fixed-rate compression, and establish a foundation for adaptive, storage-efficient learning systems on resource-limited devices. This work opens new directions toward generalizable, storage-aware on-device intelligence.