AWS aims to chip away at the machine learning barrier to entry with Amazon SageMaker, but few developers know where to start or what data to use.
Amazon SageMaker is a managed service that enables developers to build, train and deploy machine learning models. SageMaker works from data acquisition through production.
Organizations jumping on the AWS machine learning bandwagon should learn these Amazon SageMaker examples and how to get the most out of the product before they dive into any major projects.
Solve a business problem
While Amazon SageMaker comes equipped with machine learning algorithms, developers can't assume these will be right for their particular situation. Take an outside-in approach, said Charlie Kun Dai, principal analyst for enterprise architecture professionals at Forrester. Don't do machine learning just to do machine learning -- see if it can solve a business challenge, he said.
Prospective users should certainly study SageMaker documentation and tutorials, but they really need to figure out how machine learning could advance capabilities they already have. Before developers choose a machine learning model, they should understand the data they have access to, which takes time, said David Schubmehl, research director of AI software platforms at IDC.
Once an organization identifies the type of problem it must solve, developers can figure out what kind of data to use and whether SageMaker's native machine learning algorithms will suffice.
Data scientists are your friend
While Amazon SageMaker is a developer-centric platform that's run through API calls and other dev tools, not every developer knows what data will be predictive and how it should fit into a productive machine learning algorithm. Data scientists are an essential piece of the SageMaker puzzle at this point in the process, as they know how to build and evaluate a data set.
Data collection is at the heart of any AI or machine learning platform. SageMaker use depends on collaboration between developers and data scientists. SageMaker can collect data from other Amazon cloud services or in-house data repositories and store it an S3 bucket. It works with Jupyter notebooks, which data scientists and developers can use to easily share and examine training data.
Start with a chatbot
Your first SageMaker application doesn't have to reinvent machine learning. A chatbot can be a good place to start.
Organizations should use SageMaker to widen the capabilities of existing services that are already tied to AWS, said Svetlana Sicular, research vice president at Gartner. A chatbot to streamline customers' experiences or their internal operation is a good entry point as an Amazon SageMaker example.
For internal use, a chatbot could aid in compliance checks, Sicular said. Every company faces compliance challenges. Instead of having employees dive into compliance documentation to determine if they can do something, companies can build a chatbot on AWS that understands the question, finds the rules on the matter and answers it. Amazon Lex, a conversational interface service, already provides the speech recognition and natural language understanding capability, so you can integrate that capability into applications already built on AWS.
After you've created your in-house chatbot, you can then build a predictive chatbot for customers, working with both Lex and SageMaker. For this Amazon SageMaker example, a credit card company could build a chatbot that predicts whether a would-be customer qualifies for an account. Developers and data scientists would build a prediction algorithm and machine learning model on Amazon SageMaker, then configure a simple text-only chatbot with Lex and integrate both features on Lambda for serverless operation.
Evolution of Amazon SageMaker
Launched in 2017, the end-to-end machine learning service has gained features as it evolves as an Amazon cloud service, including SageMaker Ground Truth for automated data labeling and building training data sets; SageMaker Neo for faster model training; and SageMaker RL, which enables reinforcement learning capability.
AWS has also organized programs around machine and reinforcement learning, such as the DeepRacer league, where developers can test their SageMaker RL models by racing one-eighteenth scale, automated cars.
Model training has been a pricey pain point for AWS machine learning users, and the company has lowered the cost with the addition of SageMaker Managed Spot Training. SageMaker users can tap EC2 Spot Instances for machine learning model training, which AWS claims can lower these costs by up to 70%.