Computer Vision News - February 2018

24 Computer Vision News Roy Davies Nonetheless, the science of it is, to achieve these feats of recognition, you do need that much training on real data, with natural variation and real changes in lightness and darkness and contrast and so on. Neural networks have actually managed that in a fair number of cases. Face recognition is a very important case. What we’ve got to do is to try to get proper science to embody all of this. But it’s almost impossible for a scientist to define what a face ought to look like, so the only way forward is to train scientific theories on similar quantities of real data, and to identify any theoretical short-cuts that validly advance the underlying methodology. Specifically, the data has to be modelled scientifically: this is now the ‘missing link’. In summary, the value of deep learning has been to involve us in a phase of scientific discovery, but now a phase of deeper scientific understanding needs to follow. How do you compare the current generation of students with your generation? What do you think they could learn from the previous ones? Well, one of the things that has happened in recent years is that methods like MATLAB and Python have taken off. You have libraries of vision algorithms available: say edge detectors and Hough transforms … a whole range, a complete panoply of vision algorithms. So what the students know is that they have to take bits, put them together, and not think too hard about it. It’s just like using Lego. You could say that’s cheating. They’re not doing the real thing. They don’t really know what is in the blocks they are using. You need to know not only how to put the blocks together but also what is inside them to get the best out of them. Furthermore, you need to design some of the blocks in your own way in order to optimise their operation and adapt them to new tasks. In general it’s become too easy to re-use old blocks: proper training of students needs to include invention, not just assembly! Who was the teacher that impressed you the most? That’s interesting, because my father was my teacher, role model and mentor until I was about 26. However, he wanted me to be a physicist, and I wanted to be an engineer, or even an inventor. At this point I started diverging from the path he had laid down. Also, as an undergraduate, my future PhD supervisor’s lectures on electronics absolutely thrilled me. My hobby was electronics and still is for that matter. He filled in quite a lot of gaps that I didn’t know. Anyhow, in the end, I chose his research lab because it had more electronics than any other lab I could find in Oxford’s Clarendon Laboratory. There I worked at a great many frequencies between 0 Hz and 1010 Hz. All these electronics helped me Guest “I owe a huge debt to Michael Baker, my PhD Supervisor, who died in 2017” “Deep learning networks provide science with a valuable existence theorem which beckons the way forward ” Roy Davies, Frank Dellaert and Roberto Cipolla at BMVC (2013)