Companies looking to build and deploy machine learning models in the cloud have a new service from Amazon Web Services to help them. Called SageMaker, it’s designed to make it easier for everyday developers and scientists to build their own custom machine learning systems.
While machine learning can provide significant benefits to customers (assuming they have the right data), it can be hard to get started without deep expertise. SageMaker is designed to help with that by providing customers with a wide variety of pre-built development environments based on the open source Jupyter Notebook format.
Customers get pre-built notebooks for common problems, and can then pick from a set of 10 different common algorithms for a wide variety of machine learning problems, or tap into their own preferred machine learning frameworks including TensorFlow, MXNet, Caffe and others.
After that, users point SageMaker at a bunch of data in AWS’s Simple Storage Service (S3), and have it train the model. SageMaker will handle all of the work setting up data pipelines, Elastic Block Storage volumes and other components, and then tear them down when it’s done.
Users can then use SageMaker for fine tuning the performance of their models, by optimizing the hyperparameters that are baked into it. It’s a time-consuming process that is traditionally done manually, but AWS tests multiple parameter sets in parallel, and uses machine learning to optimize the process.
Once a model is trained, users can then tell SageMaker how many virtual machines they want to dedicate to running the system. It’s also capable of A/B testing models, so that users can see how their changes will affect the systems they use.