At the Deep Learning Summit in London this September, we had the fantastic opportunity to record 6 episodes of the Women in AI Podcast. We spoke with CEOs, researchers, data scientists and industry experts to hear about their research as well as their experience working in the industry.
Episode 2 of the podcast where we chat with Antonia Creswell from Imperial College is now available for you to listen to via our website or on iTunes, and in the upcoming episodes feature guests from Facebook, Maluuba, Sightline Innovation, McGill and many more. Make sure you subscribe to keep up to date on the latest episodes.
Here’s what we got up to when we spoke with Antonia:
Yaz: I’m now joined by Antonia Creswell who’s a PhD student at Imperial College. Antonia’s work focuses on unsupervised learning and generative models, and she presented at the summit earlier on her most recent research.
Thank you so much for joining us, it’s been great to have you here at the Deep Learning Summit. I was lucky enough to catch your presentation, but for everyone else listening can you give us a brief overview of your current work and research?
Antonia: Yeah! So I work primarily with generative models, so these are models where you can train them to learn from some kind of data and then reproduce new samples of that data that don’t currently exist. For instance you can show a model a whole load of faces and then you can actually generate new faces that don’t exist in the real world, like they’re not real people - you can just produce their faces.
Yaz: Okay that’s amazing! Do how did you get into this - how did you get into generative models and computer vision?
Antonia: I actually came from quite a strange background for this sort of work in some sense, I did an undergrad degree in bioengineering at Imperial, and I did a lot of projects to do with image processing and computer vision, which I decided pretty quickly, maybe even in my second year, maybe even before that, that I was really interested in this sort of thing, so I did a lot of projects in this. I did one exchange year in California and I was able to pick a lot of my modules so I picked everything I needed to do more computer vision and I did an interesting project there with a really great supervisor. For my final year project which I did in the states I actually ended up using some, we call them deep features, so it was information you can extract from a pre trained deep learning model, and basically this revolutionised my results - overnight I got way better results than the results I had, so I was immediately convinced that deep learning was the way forward for computer vision. Everyone’s kind of realising that and knows that now, but it was amazing for me to see that and then from then I decided okay I’m definitely going to go into computer vision and it might be deep learning, but then I started my PhD and I definitely don’t look back now.
Yaz: So I guess within deep learning, unsupervised learning is also something that’s growing massively at the moment, so are you using a lot of unsupervised learning in your work at the moment?
Antonia: Yeah so the great thing about generative models is that they don’t need to be trained with labelled data, so unsupervised means training without labels, so most generative models can be trained entirely without labels. Because the focus of what I do is on generative models, I pretty much focus on training without labels, which I also think as a PhD project is actually quite good because often as a student it’s quite difficult to get labelled data.
Yaz: Yep, definitely. So are you facing many challenges with this at the moment - in your work what kind of obstacles are you facing?
Antonia: The main difficulties are things with resource, I mean obviously there are a lot of technical challenges, but the main challenges at University is resource.
Yaz: Of course, that’s fair enough! So what other industries do you think need to come together to advance your work in generative models? Do you see any other disciplines being integral in helping with your work?
Antonia: Yeah absolutely definitely! With generative models, and especially with image generation, you’re trying to generate images that are perceptually coherent. For instance if you’re generating faces you want them to look real, and at the moment it’s very difficult for us to evaluate these, we don’t have any good ways to evaluate generative models except for showing them to people, and designing those perceptual experiments is not necessarily something that someone with a deep learning background is going to be very good at. You need someone with those expertise in designing these human perception experiments. I’ve tried to work with people like this and I’m trying to design these experiments, but obviously it’s very difficult if you don’t have this experience.
Keen to hear more from Antonia? We continue the discussion and touch on the highlights of the Deep Learning Summit as well as her experiences as a woman working in AI, so subscribe to the Women in AI Podcast and keep up to date with future episodes.