Computer Vision News - February 2018

Computer Vision News 23 Roy Davies 1990s. It was useful work, but the truth is that at that stage they weren’t reliable. Also, people didn’t know what these networks were ‘thinking’ or how they worked. Gradually, people moved back to conventional algorithms instead. There were statistical pattern recognition methods that could emulate the neural networks. Meanwhile, after their demise towards the end of the 90s, people had continued working with them. It was a sort of underground movement, almost invisible to many people. Suddenly, they emerged again in 2010 or so as ‘deep’ networks. By 2012-2015 they completely exceeded people’s expectations for applications such as face recognition. They actually performed better than conventional algorithms, which was a shock. I, personally, like many other scientists, didn’t like this because, again, you didn’t really know how they were working. They seemed to be working, and working well, or even superlatively. And this time you couldn’t ignore them. In those three or four years, the whole subject had changed radically. Now almost everyone wants to work with these new, deep neural networks. You are certainly aware of the article by Nikos Paragios about what he calls “ The Deep Depression ” , in which he declares that - notwithstanding the impressive performance of deep learning - the progress of science is understanding what we are doing, rather than putting layers and adjusting layers. What is your take about that? Well, I read his article, and other people think exactly the same. I almost thought the same too. The thing is that now neural networks - the new sort, the deep ones - have actually taken off in such a way that you cannot ignore them. The question is: how can science develop in that atmosphere? What we’ve got to recognise is that science advances in phases. You can go so far with one phase then you’ve got to take over with another. The thing I realized is that these neural networks were trained with millions of sample patterns, like faces from the internet. You need a million faces, or even a hundred million faces to train a network properly. But really, you must compare it with a human brain … the way children learn. They learn by playing with things and seeing everything going on all around them. They see faces, small and large, from every angle. They are learning from thousands of millions of faces in different positions. And they see the same face from different angles with thousands of millions of examples. It’s difficult for deep neural networks to get that much data autonomously. Then what happens is that the human operator has to spend his time developing methods for providing the data – typically using tricks such as chopping images into several patches. Guest Bob Fisher, Mark Nixon and Roy Davies exhibit their IAPR Fellowship awards (2008)