Microsoft Keeps Making the Right Investments in Machine Learning
A few days ago during the Ignite Conference, Microsoft unveiled a series of new tools that expand its machine learning(ML) capabilities and lower the entry point for developers and data scientists getting started with the platform. The new additions to the Azure ML stack signal that the Redmond giant is putting the right pieces in place to automate the end-to-end lifecycle of machine learning applications.
Microsoft’s entrance in the machine learning space was marked by the launch of the Azure ML cloud service. While the initial version included some impressive capabilities such as the mode designer and the ability to generate APIs from ML algorithms, it had some notable limitations when came to implementing real world machine learning scenarios. To some extent, Azure ML felt more like the type of platform that could be used by mainstream Azure developers who were not necessarily machine learning experts while “real” data scientists will rely on emerging deep learning frameworks and tools such as TensorFlow, Caffe, Theano, Jupyter, Zepellin, etc. Don’t get me wrong, Azure ML was a solid first release by Microsoft but the limitations were pretty obvious for data scientists and that has somewhat hindered its adoption. With the new release, Microsoft is directly addressing some of the well-known challenges of the initial version of Azure AML making it a more attractive destination for machine learning practitioners.
From a Runtime to a Complete Platform
The new machine learning tools released by Microsoft mark the transition of Azure ML from a runtime to an end-toend machine learning application platform. The new release include capabilities in areas such as training, deployment and management which are essential in the implementation of machine learning solutions in the real world.
Training: Azure ML Experimentation Service
The Azure ML Experimentation Service enables the rapid training and deployment of machine learning experiments. The service integrates with several deep learning frameworks such as TensorFlow, Caffe2, PyTorch, Cahiner and Microsoft Cognitive Toolkit. The Experimentation Service also enables capabilities such as model versioning or integration with source control systems.
Management: Azure ML Workbench
In Microsoft’s own words: Azure ML Workbench should become the “control panel for your development lifecycle and a great way to get starting using machine learning”. Distributed as Windows and Mac native applications the ML Workbench enables the implementation of models in Python, PySpark and Scala and integrates with Jupyter notebooks as well as IDEs such as Visual Studio Code and PyCharm.
Deployment: Azure ML Model Management
The Model Management Service packages production ready machine learning models as Docker containers that can be ported over and executed on different runtime environments. The new service capability is likely to streamline Azure ML’s integration with many advanced Docker-Kubernetes runtime stacks in the market.
Where are we at?
The release of Azure ML Experimentation, Workbench and Model Management services can become a strong differentiator with Azure’s main competitors. While services such as AWS ML or Google Cloud ML remain mostly as advanced runtimes for machine learning applications, Azure ML seems to ambition to become an end-to-end platform for the automation of machine learning applications. More importantly, the new machine learning tools represent a strong addition to Microsoft’s impressive suite of cognitive computing technologies that currently includes products such as Cognitive Services, Cognitive Toolkit, R Server, Azure ML and others.