Continuous Training and Deployment of Deep Learning Models

Abstract Deep Learning (DL) has consistently surpassed other Machine methods and achieved state-of-the-art performance in multiple cases. Several modern applications like financial recommender systems require models that are constantly updated with fresh data. The prominent approach for keeping a DL model is to trigger full retraining from scratch when enough new data available. However, large complex time-consuming compute-intensive. This makes costly, wasteful, slow. In this paper, we present an continuously train deploy models. First, enable continuous training through proactive combines samples of historical streaming Second, deployment gradient sparsification allows us send small percentage the updates per iteration. Our experimental results LeNet5 on MNIST CIFAR-10 show keeps comparable—if not superior—performance at fraction time. Combined sparsification, sparse enables very fast deployed arbitrarily sparsity, reducing communication iteration up four orders magnitude, minimal—if any—losses quality. Sparse training, however, comes price; it incurs overhead depends size increases time by factors ranging 1.25 3 our experiments. Arguably, price pay successfully enabling