Group Knowledge Transfer: Collaborative Training of Large CNNs on the Edge

Scaling up the convolutional neural network (CNN) size (e.g., width, depth, etc.) is known to effectively improve model accuracy. However, the large model size impedes training on resource-constrained edge devices. For instance, federated learning (FL) on edge devices cannot tackle large CNN training demands, even though there is a strong practical need for FL due to its privacy and confidentiality properties. To address the resource-constrained reality, we reformulate FL as a group knowledge transfer (GKT) training algorithm. GKT designs a variant of the alternating minimization approach to train small CNNs on edge nodes and periodically transfer their knowledge by knowledge distillation to a large server-side CNN. GKT consolidates several advantages in a single framework: reduced demand for edge computation, lower communication cost for large CNNs, and asynchronous training, all while maintaining model accuracy comparable to FL. To simplify the edge training, we also develop a distributed training system based on our GKT. We train CNNs designed based on ResNet-56 and ResNet-110 using three distinct datasets (CIFAR-10, CIFAR-100, and CINIC-10) and their non-IID variants. Our results show that GKT can obtain comparable or even slightly higher accuracy. More importantly, GKT makes edge training affordable. Compared to the edge training using FedAvg, GKT demands 9 to 17 times less computational power (FLOPs) on edge devices and requires 54 to 105 times fewer parameters in the edge CNN.