Back in the day when Amazon.com was just a service for books, it had already begun to explore the use of artificial intelligence (AI)-related techniques known as machine learning to automate a number of back-office functions, from recommending books to regular customers to detecting fraud. The company went on to become a machine learning expert itself in developing the use internally and has now opened up this expertise as Amazon Machine Learning service, launched on April 9, 2015. Ovum sees AI as a major growth area in the next decade, and Amazon Web Services (AWS) is now well positioned to be a port of call for developers seeking to build solutions infused with machine learning.
Amazon Machine Learning will enable developers to jumpstart their use of AI
The Amazon Machine Learning service is based on tried and tested technology that Amazon has developed internally and is designed to give a quick start for the creation of models, to visualize and optimize them, and to help put them into production on an AWS-hosted application. A key benefit is that with the large-scale models that this service supports, models can be trained with up to 100 GB of data. The service spans from the batch processing of billions of records, to the capability to make real-time predictions. The service also automatically manages the data and servers required to support the models being built and used in production. Amazon expects businesses (including rival retail stores) to use its approach.
Amazon Machine Learning is integrated across AWS data stores, Amazon S3 (storage), Redshift (data warehousing), and MySQL databases managed by Amazon Relational Data Service (RDS)). For example, to create a predictive model with data residing on these data stores, the service will automatically analyze the data regardless of its structure, make inferences, provide an interactive experience for working with machine learning models in the service console, and also allow the user to explore the data structure visually. Any malformed or missing data that breaks the user’s type validation is flagged for the user to fix. The service provides automatic or manual feature transformation. This activity usually requires some depth of data expertise so it will be interesting to see how well the automatic option works. Amazon said that by using the new service, a two-man AI project can be accelerated from a month to 20 minutes.
Once trained, a prediction model can be used by external systems to obtain predictions though a synchronous, low-latency, and high-throughput API for real-time and batch predictions. Existing AWS applications that already use Amazon event-driven services, such as Amazon DynamoDB and AWS Lambda, can link to the prediction model’s API and immediately exploit machine learning with minimal changes to the application.
There is huge potential to expand the service
The algorithms available on the service (http://aws.amazon.com/machine-learning/faqs) currently lack neural networks. Ovum suspects that for many AI engineers, this would be a prerequisite for the title of “machine learning”, so this launch should be considered as a first step. Ovum expects to see algorithms, such as deep learning, genetic algorithms, and more, to become available. Currently the focus is on prediction models involving time series and it would be useful to broaden the applications to other areas. In addition, it would be useful to be able to exploit the automated data transformation service as an independent service.
Machine Learning in Business Use Cases, IT0022-000335 (April 2015)
Michael Azoff, Principal Analyst, IT Infrastructure Solutions