|
|
|
|
||
WSJ: New Chips Propel Machine LearningNvidia microchips are helping in detection of anomalies on CT scansBy DON CLARK Computer users have long relied on Nvidia Corp.’s technology to paint virtual worlds on the screen as they gunned down videogame enemies. Now some researchers are betting it can also help save lives—of real people. Massachusetts General Hospital recently established a center in Boston that plans to use Nvidia chips to help an artificial-intelligence system spot anomalies on CT scans and other medical images, jobs now carried out by human radiologists. The project, drawing on a database of 10 billion existing images, is designed to to “train” systems to help doctors detect cancer, Alzheimer’s and other diseases earlier and more accurately. “Computers don’t get tired,” said Keith Dreyer, the center’s executive director and vice chairman of radiology at Mass General. “There is no doubt that this will change the way we practice health care, and it will clearly change it for the better.” The effort is one of many examples illustrating how advances in microchips—particularly the graphics-processing units pioneered by Nvidia—are fueling explosive growth in machine learning, a programming approach in which computers teach themselves without explicit instructions and then make decisions based on what they’ve learned. Internet giants such as Google Inc., Facebook Inc., Microsoft Corp., Twitter Inc. and Baidu Inc. are among the most active, using the chips called GPUs to let servers study vast quantities of photos, videos, audio files and posts on social media to improve functions such as search or automated photo tagging. Some auto makers are exploiting the technology to develop self-driving cars that sense their surroundings and avoid hazards. Some companies are betting that GPUs will be overtaken for such purposes by more specialized chips. Google, in a surprise move, last Wednesday disclosed that, in addition to Nvidia’s GPUs, it has been using an internally developed processor for machine learning. Others advocating special-purpose processors include Movidius, a Silicon Valley startup selling chips it calls vision processing units, and Nervana Systems, a machine learning service that plans to move from GPUs to chips of its own design. “There is no way that existing [chip] architectures will be right in the long term,” said Jeff Hawkins, co-founder of Numenta, a company started 11 years ago to work on brain-like forms of computing. For now, Nvidia has a substantial lead in the field, one of several factors that have doubled the company’s share price in 12 months and pushed its market value above $24 billion. The company, which continues to benefit from strong growth in videogames, reported this month that its business selling GPUs for data centers, rose 62% from a year earlier. CEO Jen-Hsun Huang, a Taiwan-born executive known for a trademark leather jacket and a fondness for Tesla electric cars, has emerged as a kind of Pied Piper for the machine-learning technique known as deep learning. He attributes Nvidia’s data center growth to the big cloud-computing vendors moving deep learning from testing into their core services. “It is now clear that hyperscale companies all around the world are moving into production,” he said. The research firm Tractica LLC estimates spending on GPUs as a result of deep learning projects will grow from $43.6 million in 2015 to $4.1 billion by 2024, and related software spending by enterprises will increase from $109 million to $10.4 billion over the same period. GPUs, also produced by Nvidia rival Advanced Micro Devices Inc., are especially suited for this work because they can perform many calculations simultaneously. Where conventional processors are designed to execute sequences of varied types of instructions, GPUs excel at performing a single type of calculation many times at once—like applying a color to each pixel on a computer display to generate an image. To accomplish this, Nvidia’s latest GPU has 3,584 relatively simple processor cores working in parallel, compared with one to 22 more complex calculating engines on general-purpose processors from Intel Corp. Software engineers discovered that the GPU’s massive parallel processing was especially useful in deep learning. Instead of starting with a human-made definition of a face, for example, researchers might show millions of images of faces to let a computer—sometimes with human feedback—develop its own definition of what a face looks like. The GPU can study examples much more quickly than conventional processors, dramatically accelerating the training phase. Startup Blue River Technology, a Nervana customer that has adopted GPU technology, used photographs of crops and weeds to train a camera-equipped computer system for tractors to decide where to spray herbicide. “Those machines are making 5,000 see-and-spray decisions a minute,” said Ben Chostner, the company’s vice president of business development. But some argue that GPUs simply aren’t as efficient as those designed from scratch for machine learning. Some companies, like Nervana and Movidius, emulate the parallelism of GPUs but focus on moving data more quickly and dispensing with features needed for graphics. Others, like International Business Machines Corp.with a chip dubbed TrueNorth, have developed chip designs inspired by the neurons, synapses and other features of the brain. Mr. Huang said Nvidia was well aware of Google’s development effort. He attributes Google’s motivation partly to the fact that, two years ago, Nvidia’s GPUs were better suited for training than the later phase that exploits the training to make analytical decisions. But Nvidia’s latest GPU is more than 25 times faster than its predecessor at that work, he said. |
return to message board, top of board |