On November 29th, Amazon Web Services announced SageMaker, a managed machine language service that manages the authoring, model training, and hosting of algorithms and frameworks. These capabilities can be used by themselves, or as an end-to-end production pipeline.
SageMaker is currently available with a Free tier providing 250 hours of t2.medium notebook usage, 50 hours of m4.xlarge training usage, and 125 hours of m4.xlarge hosting usage for hosting for two months. After two months or for additional hours, the service is billed per instance, storage GB, and data transfer GB.
Amalgam Insights anticipates watching the adoption of SageMaker as it solves several basic problems in machine learning.
First, this service bridges multiple gaps that businesses face in operationalizing machine learning. It is not uncommon today to see businesses use one solution to create initial models based on statistical tools, then use a second tool to build models with relevant data for training and validation purposes, and then move these models to a third solution that can effectively support these models at operational scale. The gaps in between each of these activities can lead to multi-month delays as machine learning work needs to be re-built, re-documented, or re-translated into a new environment. Amazon SageMaker will allow data scientists and analysts to use consistent frameworks, code, identities, and algorithms from initial development to production deployment.
Second, SageMaker will allow machine learning projects to focus their investment on analytic enhancement rather than basic deployment. Amalgam believes that this capability will not replace machine learning workbenches and labs, but will serve an enhancement in accelerating time-to-market. In practice, Amalgam believe that data preparation, data science workbenches and machine learning automation solutions will provide value within SageMaker, which will serve as a pipeline to translate test and experimental models into operational results. Because SageMaker focuses on consistently portraying logical, algorithmic, and destination consistency across building, training, and hosting, there is a lot of room for value enhancement within these areas such as improving data quality and relevance, optimizing model efficacy, then optimizing algorithmic delivery.
Third, SageMaker can serve as a scalable monitoring and management view for machine learning jobs across different use cases. Although job management is far from the sexiest technology role, it is currently difficult to develop a holistic view of machine learning activities within an organization because it can be difficult to simply aggregate them in a consistent manner. SageMaker will provide an easy-to-access pipeline to evaluate, train, and support operational models to serve as a potential single version of the truth for monitoring machine learning.
SageMaker is going to enhance machine learning adoption in 2018 by making it far easier for companies to get started in translating a basic model into an application-ready and infrastructure-supported tool. Amalgam highly recommends that machine learning workbenches, model automation, and data preparation solutions work both on differentiating themselves from SageMaker and figure out how they best integrate with this solution.