Over the last decade, GPU-acceleration techniques have infiltrated the high-end of supercomputing, but increased adoption of GPUs is occurring in other compute-driven disciplines too, like deep learning, one of the fastest growing segments of the machine learning field.
The trend toward accelerated computing reached a new high at last week’s ImageNet Large Scale Visual Recognition Challenge. The event, which involves the evaluation of algorithms for object detection and image classification at large scale, was the subject of a recent NVIDIA blog entry.
“When the number of users of your product flips from zero to nearly 100%, you don’t need a Ph.D. to realize a trend has formed,” states NVIDIA’s Stephen Jones.
“At last week’s event, over 95% of the teams tapped GPUs for their ground-breaking submissions. This compares with just 10% two years ago (and 0% three years ago), underscoring how accelerated computing has become fundamental for this fast-growing field,” he continued.
The strategy also reduced the team’s error rates from nearly 30 percent in 2010 to less than 10 percent in this last contest.
Winning teams, revealed at the European Conference on Computer Vision (ECCV), included the National University of Singapore, the University of Oxford, Google, the Center for Research on Intelligent Perception and Computing, and Adobe/University of Illinois at Urbana–Champaign.
The ECCV event also served as a launch pad for a new CUDA-based programming library, called cuDNN, that helps developers harness GPU acceleration. UC Berkeley researchers have integrated cuDNN into the popular deep learning framework Caffe.
In the video below Evan Shelhamer, a PhD Student Researcher at UC Berkeley explains how NVIDIA’s new deep learning software improves the behavior of Caffe.
“In Caffe, we can actually recognize the contents of a single image in only 2.5 milleseconds, which allows us to process over 40 million images a day on a single device…at massive even Internet scales,” says Shelhamer.
“With the new cuDNN library developed by NVIDIA, it’s further accelerated the key routines of deep models, so that now we can actually infer the contents of an image in just over a millisecond.
“These models can learn to do all sorts of visual tasks, even recognize the style of a photo or painting, so that it can see an image and know that it’s a vintage photo or a romantic scene or that it’s a painting done in an impressionist style.”
The graphic below shows baseline Caffe compared to Caffe accelerated by cuDNN on K40. The CPU used is a 24-core Intel E5-2679v2 CPU @ 2.4GHz running Caffe with Intel MKL 11.1.3.
More information is available here.