Democratizing Machine Learning

By Joseph Sirosh, Corporate Vice President, Information Management and Machine Learning, Microsoft

It used to be that one great technology defined an era. The steam engine, for example, served as the catalyst for the rise of the industrial age. Nowadays, however, a number of amazing technical advances and inventions are contending for bragging rights as the leading technology of our times. I would argue that one is particularly worthy of such boasting: machine learning. Although it has been in slow and steady development for years and has been used in a few enterprise applications, it has recently burst onto the scene in response to the explosion of data in today’s increasingly connected digital world. Indeed, machine learning is rapidly becoming available to the masses as a tool for making intelligent predictions on the basis of an analysis of past data.

What Is Machine Learning?

Machine learning is a branch of artificial intelligence that leverages advanced statistics. Similar to the human learning process, machine learning involves “training” computers using data labeled or classified based on previous outcomes and software algorithms that “learn” how to predict the classification of new data not labeled or classified. For example, after a period of training in which the computer is presented with spam and non-spam e-mail messages, a good machine learning program will successfully identify (i.e., predict), without human intervention, whether or not a new message is spam.

Machine learning has been applied successfully to assist in a number of specific tasks. When you swipe your credit card, machine learning algorithms consider in real time the location and amount of your purchase along with a host of other factors to determine whether it is legitimate or fraudulent. Similarly, when you shop for a product online and are presented with related choices, or with choices that might seem unrelated, machine learning is at work, comparing your purchase history with that of many other customers, some with tastes much like yours. Consumers benefit from targeted recommendations, and companies benefit from increased sales resulting from those recommendations.

Machine learning is more than the sum of its data components; it involves continual learning — inferring from the data insights that we ourselves might not grasp and then predicting possible outcomes. Critically, when you can predict outcomes you can optimize business systems and processes. For example, machine learning enables more-accurate demand forecasts, ensuring that stock levels are nearly always correct in the supply chain or even in a single retail store.

What’s New and Exciting About Machine Learning?

While machine learning has been around for a long time, until recently only people with deep skills and deep pockets could use and benefit from it. This is now changing rapidly because of both the explosion of data available for analysis and the advent of cloud-based machine learning capabilities.

There’s been a lot of big talk about big data recently, and there should be. As the volume, types, and speed of data grow, organizations of all kinds need powerful analytical models to make data-driven decisions. This requires high-performance computation that is “close” to the data and scales with the business’s needs over time. New technologies such as Hadoop facilitate the collection and analysis of large amounts of unstructured data, something we could not do before as easily — certainly not at this scale.

Speaking of scale, we are also monitoring the incredible potential of the Internet of Things, which will bring intelligence to objects and devices not only in our personal world but also in enterprises, changing the way they work. For example, Thyssen Krupp Elevators, a customer of ours, wants to know ahead of time when its elevators are ready for maintenance in order to deploy maintenance crews more effectively and minimize elevator downtime. It now relies on predictive analytics for this information, based on data provided by sensors in the elevators.

When combined with machine learning, it turns out this kind of data can help you spot malfunctions that are going to happen. Machine data analyzed by machine learning algorithms improves system reliability by determining the probability of failure well ahead of time and, eventually, making sure most things that matter never fail.

Realizing a New Vision for Machine Learning

Since I was 12 years old, I have wanted to work with artificial intelligence. I did a PhD in neural networks, and my first job involved creating a machine learning application for credit card fraud detection. At Amazon, I spent nine years in a variety of roles, but I most enjoyed building and running the company’s central machine learning group and being responsible for risk management for Amazon and its subsidiaries. I came to Microsoft in 2013 because the CEO gave me the opportunity to realize a new vision for making machine learning a mainstream application that can be used by all, leveraging the cloud.

We have a vision of creating the Intelligent Cloud, which would make machine learning more accessible to every enterprise and, over time, every one of us. Machine learning software today is usually managed on premises by each organization using it, and building machine learning applications requires expert data scientists. However, data scientists are in short supply, commercial software licenses can be expensive, and popular programming languages for statistical computing have a steep learning curve. Even if a business could overcome these hurdles, deploying new machine learning models in production often requires months of engineering investment. Scaling, managing, and monitoring these production systems require the capabilities of a very sophisticated engineering organization, which few enterprises have today.

Microsoft’s Azure Machine Learning is helping organizations meet those challenges. Combined with the rest of Microsoft’s data platform, it allows our customers to create entirely new solutions that deliver on the promise of predictive analytics dramatically faster and cheaper, and is leading them to trustworthy and actionable business insights on big data.

For example, eSmart Systems of Norway is pioneering smart grid management using our tools. A traditional smart grid — an electricity supply network that uses digital communications to detect and react to local changes in usage — includes multiple data silos, including SCADA networks, building automation systems, and substation meters. In this environment, it can be difficult to forecast consumption and prevent bottlenecks or outages. For a utility company, upgrading its entire infrastructure would be costly. Even when upgrades are made, with new smart sensors or meters, data often gets collected but is not readily accessible. eSmart Systems is now using our cloud platform to integrate and analyze usage data and create forecasts. Azure Machine Learning is the “brains” of their solution, running the data models for predictive analytics. The analytics are used to predict capacity problems and control load automatically in individual buildings.

Mendeley is another innovative customer. One of the biggest repositories of scientific research content in the world, Mendeley provides a global platform and social network to foster discovery and community collaboration. To improve the user experience, Mendeley was looking to anticipate the behavior of new users in their initial adoption and engagement phase. Within two weeks of implementing Azure Machine Learning, developers were able to create a predictive model that was 30 percent more accurate than an earlier model that had taken them months to develop on their own. Not only is Mendeley able to iterate and deploy models three to five times faster, it can pinpoint users’ needs with much greater confidence.

Building the Data Science Economy

We’re beginning to see what happens when we make machine learning accessible to enterprises. But what about making machine learning in the cloud available to individuals? I’m thinking here mostly about the supply side of the market for advanced analytics.

The cloud as a platform has already given us the app economy, where app creators reach consumers directly. There is no middle man that tells app developers what to build. They publish to an online marketplace and consumers select what’s best or most relevant for their needs.

We are trying to emulate this by adding machine learning models and packages to the Azure Marketplace: a marketplace where data scientists can show their creativity and monetize it. By “data scientists” I mean engineers and physicists and statisticians and business school graduates who love data, are passionate about developing analytical models, but haven’t had the tools to build full solutions.

With a ready-made marketplace to showcase their skills, data scientists can develop innovative analytical models, package them into APIs (application program interfaces) that others can consume, and publish these APIs. Developers and consumers can then access the same marketplace, search or browse for APIs, and pay for a specific API they wish to consume — something that they would, in turn, use to deliver a predictive analytics solution that would improve their work or lives. Examples of APIs offered in the marketplace today include those offering recommendations for similar products purchased, detecting anomalies in data, and performing sentiment analysis on textual data such as social media feeds or web pages. These marketplace APIs can be consumed in other applications or even in Excel spreadsheets. They support transactions in many currencies and offer an efficient platform for building a data science economy. We expect eventually to host millions of such analytics APIs in the cloud.

Machine learning is poised to be a game changer across industries and an important technology for improving our personal lives. We are rapidly realizing the vision of democratizing the use of machine learning by both enterprises and individuals. Businesses should take the time now to understand the true potential of machine learning and advanced analytics within their own organizations.

Originally published in CTO Straight Talk #2