SIEMENS

They work in power plants, factories and hospitals. You can find them in the basements of large buildings, in surveillance centers, and postal automation facilities. Some dart back and forth relentlessly as they drill through metal parts. Some sit stolidly in huge halls producing only a mantric hum — and lots of electricity. All are members of a new generation of systems: machines that learn.

Whether the goal is to make interventional cardiac procedures safer by automatically identifying key anatomies, or to increase the efficiency of the world’s largest turbines, the ability to learn from experience is transforming machines into systems that remember, evolve, and sometimes surprise us. Call it a paradigm shift or a revolution in the making, machine learning is set to accelerate knowledge acquisition across industries and become a decisive competitive factor as we zoom up the “on” ramp to intelligent systems.

“Learning is the gateway to intelligence,” says Prof. Tomaso Poggio, Eugene McDermott Professor in the Department of Brain and Cognitive Sciences at the Massachusetts Institute of Technology’s Artificial Intelligence Laboratory (see article “Gateway to Intelligence”). “As the complexity and specialization of our civilization continues to increase, machine learning will become the best way of sifting through all the data we are generating and identifying what can be optimized. Already, when it comes to highly complex processes, it is the only solution.”

Representing the World Mathematically. And what could be more complex than predicting the prices of electricity or copper, which are themselves based on thousands of variables? Yet that’s exactly the kind of thing — among many examples — that Siemens’ Software Environment for Neural Networks (SENN) learning system has been designed to do (see article “The Science of Prediction”). And it works! Siemens already gets decision support for its electricity purchases in Germany and its huge worldwide copper purchases from predictions provided by this unique software. “The science of prediction,” says Siemens Senior Principal Research Scientist Dr. Hans-Georg Zimmermann, who holds most of the patents that make SENN possible, “is a race between the increasing complexity of the real world and our accelerating ability to mathematically represent it by means of information technology-related capabilities.”

A good example of the relationship between complexity and IT solutions is our growing capability to replace open heart surgery with “interventional” procedures that can be performed by means of a catheter. Such procedures, however, make it necessary to practically be able to see through a patient. To meet this requirement, scientists are developing systems trained on thousands of images of human hearts. Such systems have learned to extract the outlines of, say, an aortic valve from angiography and ultrasound images, discover the anatomical landmarks that are common to both, and combine these images to produce a single hybrid view (see article “Body of Knowledge”).

Unlike human learning, which becomes less and less capable of extracting any useful message from data as the number of inputs increases, artificial learning systems positively thrive on such complexity. For instance, learning systems have produced surprising insights in experiments involving the detection of possible relationships between different parts of the human genome — an area no human brain could possibly process (see article “Thriving on Mountains of Data”). “In fact,” says Prof. Bernhard Schölkopf, Director of the Max-Planck Institute for Intelligent Systems in Tübingen and Stuttgart, “once software has been trained in this area, the more data you give it, the more precise the results become.”

That’s also true of the relationship between local weather conditions and a wind farm’s output, which is characterized by huge amounts of data that must be processed in real time (see article “On-the-Job Optimization”). For instance, Siemens researchers have developed an autonomous learning system that collects sensor data on local conditions such as wind speed, turbulence, temperature and pressure and uses algorithms to correlate this data with the farm’s output. The software gradually learns the interrelationships between inputs and outputs and adjusts variables such as the rotor blades’ angle of incidence. Over time — and as it assembles more and more experience from data — the system can produce significant improvements in a farm’s collective output.

Wind farms may also benefit from machine learning systems when it comes to external maintenance. For instance, after a major storm, an operator may wish to have its masts and propellers checked for damage — a job best performed by means of close visual inspection rather than with binoculars. Solution? How about calling in a fleet of flying robots? With this in mind, researchers at Siemens Corporate Technology in Princeton and the Massachusetts Institute of Technology in Boston are developing a “quadcopter” — a small, flying video platform that uses lasers and optical sensors to create 3D models of its surroundings (see article “Flying Inspectors”). The device, which has been test flown to inspect huge industrial facilities and produce detailed 3D digital maps of internal environments to support major upgrades, could be trained to detect and map the locations of damaged areas in wind farms.

What do envelopes, license plates, road signs, pharmaceutical products, and supermarket shelves have in common? Three things: letters, numbers, and the need for systems that can automatically read their content. And the key to meeting all of these challenges is machine learning (see article “We Read You Loud and Clear”). The essential technology behind Siemens’ world-leading position in address-reading systems for postal distribution centers, machine learning not only has made it possible to read up to 95 percent of all texts (including those that are handwritten) without error, but is now being used to help cities such as London enforce road pricing through smarter and smarter license plate reading systems — a potentially huge worldwide market. And in the security arena, learning-based machine-reading systems are being explored by Germany’s Federal Ministry of Education and Research as part of a system for tracking trucks that would be specially labeled when carrying hazardous materials.

Even machine tools, such as the heavy-duty drills and lathes used in factories, are rumbling into the learning systems marketplace. For instance, in a program initiated in 2008 at Siemens’ Technology-to-Business Center in Berkeley, California under the direction of Dr. Sarah Peach and recently transferred to Siemens Corporate Technology in Princeton, New Jersey for further development, Dr. Linxia Liao and Zack Edmonson are now working with the company’s Motion Control business unit to put the finishing touches on software called “Plug and Prognose” (PnP). The software allows machine tools to learn continuously from sensor inputs such as vibrations, current, torque, speed and temperature, and adjust their output accordingly to meet optimized values. This makes it unnecessary for the machines to go offline for testing by a specialized technician. The software also takes the need for production line flexibility into account. “For instance,” says Liao, “When a new order comes in that requires drilling through thicker slabs of aluminum, PnP communicates with a machine’s controller and adjusts associated algorithm parameters accordingly. The PnP algorithm automatically adapts to the change without requiring any user intervention. In short, it learns from experience.”

All in all, from learning to decipher the content of medical images to instantly reading license plates and envelopes, and from identifying potential maintenance problems to predicting the future, machine learning can be an accelerator for just about any technology. Nevertheless, there are still some very basic things that it cannot do. Take the simple task of figuring out what’s happening in a photograph of people at a party, for instance. “I think that would be one of the most intellectually challenging things for a machine to do,” says MIT’s Poggio. “We now have systems like Watson that can answer complex questions. We have systems that count the number of people or cars in an image. But telling what’s actually happening in an image? I think it will take at least twenty years before any artificial system will be able to do that.”