A team led by a Microsoft researcher has won the prestigious Marr prize in computer vision research for a paper that presents an alternative to what is currently the most popular method of teaching computers to recognize images.
The Marr prize winner was announced Monday at the International Conference on Computing Vision in Santiago, Chile. The prize is awarded every two years and is considered a top honor for computer vision researchers.
The award-winning paper, Deep Neural Decision Forests, showed that by using a technique called random forests researchers could create a system that was just as good, if not better, at teaching computers to recognize images than the system many leading computer vision researchers are using today. In addition to Criminisi, the authors of the paper are Peter Kontschieder, Madalina Fiterau and Samuel Rota Bulo.
These image recognition breakthroughs could eventually be used for all sorts of practical applications, such as searching through thousands of photos for ones containing just one particular person, or helping build tools that can have true artificial intelligence to speak, see and even understand.
Random forests and neural networks
A random forest is a set of decision trees, or mathematical structures that are used to organize data and make predictions.
The prevailing technique for image recognition uses convolutional neural networks, which are inspired by the biological processes of the brain.
In recent years, researchers have made huge leaps in computer vision accuracy using neural networks, creating systems that can recognize images with startling accuracy and even do things like answer questions about what's in a photo. Neural networks also have been key to major leaps in speech recognition and translation advances.
Antonio Criminisi, a principal researcher in machine intelligence and perception with Microsoft Research Cambridge in the United Kingdom, said the sudden popularity of the neural networks research got the researchers thinking about whether researchers were in danger of thinking too narrowly, and of not opening themselves up to alternative possibilities.
"In research, it's always a bit worrying when you have many, many research labs focusing on either the same solutions or very small variants around the same solution," he said. "It feels a little bit like creativity gets a hit."
Criminisi likened it to the importance of making sure that your company includes people with all different kinds of backgrounds and perspectives.
"You want to encourage diversity in the workplace," he said. "You also want to encourage diversity in research."
The award-winning paper, written by Criminisi and a team of researchers, showed that by incorporating some of the properties of convolutional neural networks into a method using random forests they could in some instances beat the best competing system that was available at the time the paper was written.
Christopher Bishop, laboratory director of Microsoft Research Cambridge, said the research is very significant because it shows how to combine the latest advances in neural networks with an older, more established approach that uses computing resources more efficiently.
"This prize-winning paper unites these two technologies in a single architecture, creating an opportunity to deliver high accuracy at low computational cost," he said.
The Marr prize award comes just one week after another team of Microsoft researchers took first place in several major categories of the ImageNet and Microsoft Common Objects in Context (MS COCO) image recognition challenges.
That winning team's research relied on an extremely deep neural network, which is as much as five times deeper than any method previously used.
Criminisi said his team is now talking to the team that won the ImageNet and MS COCO challenges about ways the groups can work together to find even better methods for recognizing images. The hope is that by collaborating they can come up with solutions that are even better together
No comments:
Post a Comment