On-Device Dataset Distillation: The MNIST Use Case and Initial Experiments on Agri-Food product classification

The agri-food sector generates large volumes of data from sensors monitoring environmental factors such as humidity and temperature, which impact product quality and safety. However, resource constraints make real-time processing and storage challenging. This study explores dataset distillation as a data reduction technique to optimize data efficiency. We evaluate this approach using MNIST, a standard benchmark dataset, with reduction ratios of 66.67% and 90%. The distillation process was implemented on an M5StickCPlus (M5C+), with data exchanged via MQTT between the device and a Raspberry Pi. The model trained on distilled data was tested on the original dataset, demonstrating strong generalization despite significant data reduction. Beyond MNIST, we conduct a preliminary experiment applying dataset distillation to a pistachio image dataset, illustrating its potential for agri-food image classification applications. The results indicate that distillation can effectively retain essential visual features while reducing data redundancy, making it a promising technique for resource-constrained environments. In Federated Learning, it can aggregate multiple samples while retaining only a distilled subset for on-device training, reducing storage and computational requirements. In centralized learning, it minimizes energy usage and optimizes bandwidth by transmitting only distilled data instead of full datasets, lowering transmission overhead and preserving device storage capacity. These findings highlight the potential of dataset distillation for enabling efficient and privacy-preserving machine learning in both general and agri-food applications.