ImageNet Challenging Classification with the Raspberry Pis: A Federated Learning Algorithm of Local Stochastic Gradient Descent Models

Abstract

In this paper, we propose the federated learning algorithm of local stochastic gradient descent (SGD) models at edge devices (i.e. Raspberry Pis) to classify large ImageNet dataset having 1,281,167 images with 1,000 classes. The full very large training dataset is divided into subsets which are stored in local Raspberry Pis. And then, the federated learning algorithm uses Raspberry Pis to train in the incremental and parallel way local SGD models from their own subset without exchanging data. The incremental local SGD tailored on Raspberry Pi sequentially loads small data blocks of its own local training subset to learn local SGD models. In which, the local SGD algorithm uses kmeans to split the data block into k partitions and then it learns in the parallel way SGD models in each data partition to classify the data locally. The numerical test results on Imagenet dataset show that our federated learning algorithm of local SGD models with 4 Raspberry Pis (Broadcom BCM2711, Quad core Cortex-A72 (ARM v8) 64-bit SoC @ 1.5 GHz, 4 GB RAM) is faster than the state-of-the-art linear SVM run on a PC (Intel(R) Core i7–4790 CPU, 3.6 GHz, 4 cores, 32 GB RAM) with the competitive classification accuracy.