Computer Vision News - December 2016

Nikos Paragios is full professor of Applied Mathematics at the Ecole Centrale Paris and the scientific leader of an INRIA research group called GALEN at Paris-Saclay . He is also Editor in Chief of the Computer Vision and Image Understanding Journal . He wrote some time ago a very intriguing text about deep learning called Computer Vision Research: The deep "depression" . He was so kind to discuss with Computer Vision News about scientific research, education, robots, industry and more. He also gives very precious advise to all students… Computer Vision News: Nikos, I know that some time ago you wrote a very original text about deep learning. Can you share with our readers your intriguing views on deep learning? Nikos Paragios: Deep learning is the hottest topic in computer vision and beyond in our days. We have more and more people jumping in, and the reason why this is happening is because the performance is impressive. We’re getting results that were unimaginable with state of the art methods five or ten years back. The reason why this is happening is because we have more data and computing power. We also have methods that were developed in this area, like neural nets , that have been around for 40 years and have progressed in terms of how you can optimize and learn these networks. I wouldn’t say I’m not fond of it, but the reason why something is bothering me is because I have the impression that the community is actually moving forward building architectures: it’s becoming a way of composing a puzzle without having a clear picture of what is happening behind. I think something like 50% of the community is working in this area. Some people are doing great theory stuff. But the rest of the community is actually trying to get the results out of these methods by decomposing, putting layers and adjusting layers. I think that shouldn’t be the main objective of academic research. This is something that should happen in industry. Our job is to progress science and the progress of science is understanding what we are doing. So I think it’s a great tool, its performance is really impressive and it has all the potential to become a very advanced tool in the field. But I prefer that we spend more time on theoretical questions rather than spend time trying to beat results on a benchmark. CVN: It does help us create better applications in real life... Paragios: Yes, it’s already happening. And this is one of the reasons why deep learning is very popular: because even without having a very strong expertise in the area, if you have a dataset you can build very good prediction and recommendation systems. That’s the reason for its success, it is a great tool and it contributes to make computer vision so popular in the industry. If you go to conferences now, the industry presence with its demos is as strong as posters. I remember 5 years ago there were only a few companies. Now there are more and more small, medium, and large sized companies in different areas. It’s actually helping us a lot to transfer technology because it works. 4 Computer Vision News Guest Professor Nikos Paragios Guest “ Our job is to progress science and the progress of science is understanding what we are doing”