Cloud-based Collaborative Agricultural Learning with Flexible Model Size and Adaptive Batch Number

Abstract

With the rapid growth in the world population, developing agricultural technologies has been an urgent need. Sensor networks have been widely used to monitor and manage agricultural status. Moreover, Artificial Intelligence (AI) techniques are adopted for their high accuracy to enable the analysis of massive data collected through the sensor network. The datasets on the devices of agricultural applications usually need to be completed and bigger, which limits the performance of AI algorithms. Thus, researchers turn to Collaborative Learning (CL) to utilize the data on multiple devices to train a global model privately. However, current CL frameworks for agricultural applications suffer from three problems: data heterogeneity, system heterogeneity, and communication overhead. In this paper, we propose cloud-based Collaborative Agricultural Learning with Flexible model size and Adaptive batch number (CALFA) to improve the efficiency and applicability of the training process while maintaining its effectiveness. CALFA contains three modules. The Classification Pyramid allows the devices to use different sizes of models during training and enables the classification of different object sizes. Adaptive Aggregation modifies the aggregation weights to maintain the convergence speed and accuracy. Adaptive Adjustment modifies the training batch numbers to mitigate the communication overhead. The experimental results illustrate that CALFA outperforms other SOTA CL frameworks by reducing up to 75% communication overhead with nearly no accuracy loss. Also, CALFA enables training on more devices by reducing the model size.