Posted by Rob Knies

Andrew Blake during the Machine Learning Summit 2013

Unlocking the future—that was the theme Rick Rashid, Microsoft chief research officer, used to close his opening remarks April 23 during the first day of the Microsoft Research Machine Learning Summit 2013.

The event, held at Microsoft’s Le Campus site in Issy-les-Moulineaux, France, just outside of Paris, gathered thought leaders and researchers from a broad range of computing-related disciplines to focus on key challenges in a new era of machine learning and to identify what will be necessary to take advantage of the information resources of today and tomorrow to enhance society at large.

Co-chair Evelyne Viegas of Microsoft Research Connections opened the summit with a few introductory remarks before introducing Alain Crozier, president of Microsoft France, who welcomed the approximately 250 attendees to the event. Viegas then took the opportunity to bring Rashid to the stage for his introductory remarks.

“This topic of machine learning has become incredibly exciting over the last 10 years,” he said. “The pace of change has been really dramatic, so it’s exciting to get so many people from so many different areas to be here today to talk about it.”

Rashid recalled a time in the not-so-recent past when machine learning was in its nascent stage, using rules and pattern recognition to produce results many found less than stellar. But today, he stated, a combination of data, devices, and services has led to newfound respect for the discipline, which is having an increasingly dramatic impact on business and society.

Rashid discussed advances in deep neural networks that have demonstrated dramatic improvements in speech-to-speech translation, tracing a path from speech recognition to machine translation to personalized speech synthesis. He shared with the audience a video of his rapturously received talk in Beijing in October in which his remarks in English were transformed into Mandarin, in his own voice, before a delighted group of Chinese academics.

“I think what’s interesting,” Rashid concluded, “both about this demonstration, but also about the changes that are now happening in this field of machine learning, especially as we bring new techniques like these deep neural networks, is the sudden optimism that for some of the great problems of computer science … now, we may actually be beginning to see the light at the end of the tunnel, that we may see, not just in our lifetime but maybe in not very many years, many of those critical problems be solved.”

With that, Christopher Bishop, a Microsoft distinguished scientist based at Microsoft Research Cambridge and the other summit co-chair, summoned Andrew Blake, laboratory director at that facility, to discuss computer vision in a keynote talk called Machines That (Learn to) See that will be streamed live, beginning at 1330 GMT.

“In studying computer vision,” Blake said, “we have an immense, privileged view of the study of intelligence. There are many scientists trying to study intelligence from different points of view. Some of them have set themselves the task of understanding in detail what is the neural circuitry of the brain. Accepting those constraints of a fully biological explanation of intelligence is a tremendously heavy burden.

“What we have in computational studies is the opportunity to do this from a completely different perspective, where we study intelligence, in some sense, in an abstract viewpoint, unburdened by some of the details.”

Blake went on to describe the two modern-day styles of computer-vision research, that of empirical detectors and that of analysis by synthesis. He left little doubt that he was particularly captivated by the latter, mentioning advances in the development of the Kinect human pose-estimation pipeline and the Photosynth system of amassing 3-D image reconstruction from a collection of photos of the same subject.

He pointed toward a future, though, where the two styles would be fused to produce even more advances, suggesting that data-driven sampling might be one promising step forward.

“While we think of the problems and how to solve them in the most systematic way,” Blake concluded, “there are so many things in vision now that really work, whether it’s commodity cameras, commodity software, medical imaging, cars that drive themselves, and new user interfaces. It’s a wonderful achievement that vision has become so mature.”