ICCV Daily 2019 - Thursday

Ruth Fong is a PhD student in the Visual Geometry Group (VGG) at the University of Oxford, where she is advised by Professor Andrea Vedaldi and funded by the Rhodes Trust and Open Philanthropy. Where are you from originally? I’m originally from the US, from New Jersey. You have presented an oral and a poster here, yesterday. Can you tell us about your work? The broad research topic that I’m interested in is interpretability, which is just trying to understand what these models are actually doing. Especially because the state-of-the-art right now is the deep neural network systems, this is particularly important. These are black boxes that we don’t understand. The work that we presented yesterday was really trying to tackle this problem of attribution, which is trying to understand what part of the input, in this case because we deal with images, what image regions are responsible for the model’s prediction. We’re trying to generate a heat map that explains when a dog was predicted for this image, whether it was looking at a watermark, whether it was reading the word “dog” in the image, or whether it was looking at the ears, the nose, or something else. Why did you choose that field? I’m not quite sure. I think a lot of things are quite stochastic. I’m not sure quite how I ended up in computer vision, but my first taste of research was in David Cox’s lab back at Harvard where I did my undergrad. It was part computer science, part neuroscience lab with a focus on vision. One really big motivating factor for why I chose vision was I wanted something that was really easy to explain to anyone. That’s how I ended up in vision. The reason why I am interested in interpretability is because, in this century, it looks like AI is going to be the dominant force behind the technological revolution. A lot of these big ethical questions need to be answered. Interpretability helps us develop techniques or justbetter understand, both from a scientific perspective, what we are doing. I originally thought I was going to be a biologist. I like that interpretability. It’s kind of like you’re asking scientists questions about the model, and you’re 14 Women in C. Vision DA I L Y