The two American giants Google and Nvidia peak at the top of artificial intelligence training programs. At least that’s the opinion of MLCommons, the industry consortium that oversees a popular machine learning performance test, MLPerf.
The latest results of this test reveal that Google is the best performing company in terms of time to train specialized neural networks in image recognition, object detection, testing of small and large images and for processing BERT natural language.
For its part, Nvidia performs with regard to image segmentation, voice recognition, recommendation systems and resolution of the reinforcement learning task of the Go game on the “mini Go” data set. Benchmark tests deployed as part of MLPerf consider the number of minutes it takes to tune neural parameters until the computer program achieves a minimum precision required for a given task – in a process commonly referred to as “training” a neural network.
Moore’s Law outdated?
Google and Nvidia are not the only players in machine learning and artificial intelligence to show clear progress in this area. Across all vendors, training times have improved dramatically due to greater power and smarter software approaches.
For David Kanter, executive director of MLCommons, the results of the latest edition of MLPerf show that training performance has increased faster than Moore’s Law, the traditional rule that doubling the number of transistors in a chip every 18-24 months speeds up computer performance.
Scores from the venerable ImageNet task, in which a neural network is trained to assign a classification label to millions of images, are 9 to 10 times faster today than what simple chip improvements would imply, notes on this David Kanter. “We have done much better than Moore’s law,” rejoices the latter. “You would expect to get about three and a half times better performance, assuming the transistors have a linear relationship to performance; in fact, we get 10 times Moore’s Law.”
Source: ZDNet.com