Machine Learning - Research (EN)

NucleusResearch.com 8 Document Number: T147 October 2019 The company used SageMaker to deploy the model responsible for predicting player attrition into production. It was chosen for because it gives the project leadership a "birds' eye view" of the model and affords project leadership a centralized location to view and control all the models in the deployment. With technology like SageMaker combined with the flexible compute resources offered by Amazon, the interviewed expert said, "it would take a compelling business case to persuade us to leave AWS and so far no other cloud provider has been able to offer us the machine learning-specific tools along with storage and compute for a [better value]." AWS INVESTMENT IN DEEP LEARNING Customers know that Amazon is developing and using its own deep learning technology. Deep learning experts referenced the ongoing improvements to documentation, framework support, and cloud services like Amazon SageMaker as primary factors in choosing AWS over other cloud providers. SageMaker became available in 2017; it is a fully-managed cloud service that covers the entire machine learning workflow - from building, training and deploying machine learning models. SageMaker adoption is growing fast as developers realize the how it can reduce complexity and accelerate model deployment. Last year, approximately one third of the respondents were using or exploring the use of SageMaker to automate aspects of their deep learning projects; this year that figure nearly doubled with 63 percent of customers using or considering SageMaker. In the course of this study, we found customers who were migrating their homegrown TensorFlow deployments to a managed service on AWS via SageMaker, as well as customers who built their systems from the ground up entirely with SageMaker. Users said: • "We don't need to source dedicated hardware to run large-scale deep learning projects. Without AWS and specifically SageMaker, we would need to buy hardware, train the model locally, then store and host the model on an internal server so it would be accessible when I want it for forecasting. Just to get started, this could take weeks and comes with a ton of added costs for hardware, electricity, and personnel, not to mention all the lost time from building capabilities internally that are pre-built on SageMaker." • "We focus on implementing conversational machine learning for our clients and SageMaker streamlines model building. On average we have five projects in progress for a client on a given week. Since starting to use SageMaker last year, we found that we save about two hours per project from automating distributed model training."

• Cover