Xiaofang Zhao

Wenting Tan, Xiao Shi1, Cunchi Lv et al.

Geo-distributed ML training can benefit many emerging ML scenarios (e.g., large model training, federated learning) with multi-regional cloud resources and wide area network. However, its efficiency is limited due to 2 challenges. First, efficient elastic scheduling of multi-regional cloud resources is usually missing, affecting resource utilization and performance of training. Second, training communication on WAN is still the main overhead, easily subjected to low bandwidth and high fluctuations of WAN. In this paper, we propose a framework, Cloudless-Training, to realize efficient PS-based geo-distributed ML training in 3 aspects. First, it uses a two-layer architecture with control and physical training planes to support elastic scheduling and communication for multi-regional clouds in a serverless maner.Second, it provides an elastic scheduling strategy that can deploy training workflows adaptively according to the heterogeneity of available cloud resources and distribution of pre-existing training datasets. Third, it provides 2 new synchronization strategies for training partitions among clouds, including asynchronous SGD with gradient accumulation (ASGD-GA) and inter-PS model averaging (MA). It is implemented with OpenFaaS and evaluated on Tencent Cloud. Experiments show that Cloudless-Training can support general ML training in a geo-distributed way, greatly improve resource utilization (e.g., 9.2%-24.0% training cost reduction) and synchronization efficiency (e.g., 1.7x training speedup over baseline at most) with model correctness guarantees.

Hui Zhu, Xiaofang Zhao

Dropout regularization has been widely used in deep learning but performs less effective for convolutional neural networks since the spatially correlated features allow dropped information to still flow through the networks. Some structured forms of dropout have been proposed to address this but prone to result in over or under regularization as features are dropped randomly. In this paper, we propose a targeted regularization method named TargetDrop which incorporates the attention mechanism to drop the discriminative feature units. Specifically, it masks out the target regions of the feature maps corresponding to the target channels. Experimental results compared with the other methods or applied for different networks demonstrate the regularization effect of our method.