Azure Machine Learning 2015
Microsoft seems to be serious about machine learning, having acquired Revolution Analytics (the commercial version of R), and now launching Azure Machine Learning. The company seems to have more urgency now it has a new CEO and certainly cannot be accused of dragging its feet.
And so to Microsoft Azure Machine Learning, a cloud based machine learning (ML) platform with a marketplace for models, and a toolkit for developing them. This is becoming a popular paradigm with vendors such as FICO and Algorithmia (recently done a deal with Dell), and is a pointer to the future – the real tech-heads developing stuff and app developers using it. Microsoft is also offering free access with a €150 credit to spend on the full unlimited facility, and a free subscription that does not require any kind of buy-in, but is limited in resources. After that, pricing is low for basic subscription and use of ML Studio, the model development environment. Production usage of models is where Microsoft will make its money, but even here it’s not particularly expensive at €1.49 per hour of CPU.
The Azure Marketplace allows users to make their virtual machine available to relevant markets, and models developed using Azure ML can be made generally available. It also provides a variety of data sources – useful in many machine learning applications. A host of ML applications already exist including market basket analysis, anomaly detection, text analytics, customer churn – and many more built with Azure ML.
Azure Machine Learning Studio is loaded with a large number of algorithms for the usual tasks of regression, clustering, classification and anomaly detection. These include SVM, PCA, logistic regression, neural network, multiple other types of regression (Poisson,quantile, linear, Bayesian), random forest – and so on. R and Python programming is also supported through Execute R Script and the embedding of Python scripts. A variety of data preparation steps are supported (as they need to be), and the support for Hadoop and SQL Server databases should satisfy most needs. Machine Learning Studio has a graphical interface that runs in the browser with animation and clear display of workflows.
Microsoft has not skimped on its Azure ML platform, and both model producers and consumers will find it a delight to work with. Various commentators have praised its functionality and ease of use – one such test drive of the platform was reported here. I think the competition (Amazon and IBM’s Watson) should be very afraid.