It was only a matter of time. What with Google, IBM, and Microsoft now offering cloud-based machine learning services of various kinds, Amazon's been obliged to step up with its own such product.
Amazon's just-unveiled service, Amazon Machine Learning (or AML for short), offers simplified ways to make both batch- and real-time predictions from data stored in Amazon. But it comes with some limitations.
AML works with data stored in Amazon S3, Redshift, or RDS, and provides an API set for creating, connecting with, and manipulating data sources, models, predictions, and evaluations. Data sets imported into AML can be explored visually as a way to spot-check them for inconsistency -- a feature akin to part of what IBM Watson Analytics offers. Three basic kinds of analysis -- binary classification, multiclass classification, and regression -- can be run against the provided data to produce a model. Said model can then be used in turn to classify data uploaded from a data set or fed in real-time.
Based on a walkthrough of the steps to perform a simple analysis, as shown on the AWS blog, loading data and performing interactions with it (e.g., adjusting the metrics used for a prediction) is like following an automated wizard. Pricing for the service is based on the time needed to analyze data and to build models, as well as by how many predictions are requested (with real-time predictions at a premium). The only code that has to be written is a data-transformation recipe. AML makes a best guess at concocting one, but users can edit it if need be.
The downside of this guided approach is that it's also closed-ended. Models, once created, can't be exported from or imported back into the service, according to the service's FAQ. To that end, any work done with AML ends up being tied to Amazon. Training datasets are also limited to 100 GB in size.
Microsoft's offering, by contrast, may be harder to get started with, but it offers a broad range of algorithmic modules, and allows custom R or Python code to be used as part of any analysis. On the other hand, Azure ML training data sets can only be a maximum of 10 GB in size, but multiple data sets can be combined for training. (Amazon doesn't appear to offer a function like this.)
If Amazon is taking a narrower approach than Microsoft, it's most likely because Amazon is pitching the service to developers who don't want to submerge themselves in the minutiae of creating ML-powered applications. An outfit already leveraging Amazon's cloud that wants to quickly develop a fraud detection or recommendation engine ought to find AML a boon, but the service will need to find ways to grow as its users' ambitions also ramp up.