- Machine learning has potential to improve lives and advance society
- Current state of machine learning is more art than science
- Automated methods could democratize access to machine learning solutions
The ISC High Performance 2018 conference held in Frankfurt, Germany from June 24-28 will be the biggest event in high-performance computing (HPC) in Europe this year. This year, one whole day (Wednesday, June 27) will be devoted to the increasingly vital topic of machine learning.
Science Node recently caught up with Frank Hutter, professor of machine learning at the University of Freiburg, and ISC 2018 keynote speaker, to find out more about why machine learning has become such a hot topic in HPC.
What makes machine learning so important right now?
You can do almost anything with machine learning these days. The exciting thing about doing methodological work in machine learning is that once you've figured out how to improve, for example, the way we build deep neural networks, then this directly translates into better results for a very broad range of problems.
You can think of all kinds of applications. We all know speech recognition on our smart phones, autonomous cars need to detect other cars and pedestrians from their cameras and other sensor information, and the list goes on. There are weather forecasts, or forecasts of what may happen next with Hawaii’s volcanoes, where to evacuate people, things like that. In principle, you can also use similar technology to predict whether public policy making is going in the right direction: ‘If I enact this law, what will happen in the future?’
One application I really like is in neuroscience: machine learning can be used to interpret signals coming from the brain, and this can lead to effective brain-computer interfaces. With EEG decoding you can, to a certain extent, read people's thoughts; for example, if a patient is paralyzed and can't move their limbs—can't even move their eyes in the worst cases—you could use machine learning to help them communicate. For example, tell from their EEG signals whether they are thirsty or hungry. Very basic things, but it would give them back the opportunity to interact with the world and help them achieve a basic standard of living.
Machine learning is also related to optimization. Whenever I see processes that are clearly inefficient, that kind of irks me. For example, when I take public transport and I know there would be a better way of scheduling routes so that people have to wait less or so that the system is more robust when a tram doesn't run. Machine learning is key to making improvements in such areas, because once you can predict what will happen if you make a certain change, then you can make the changes that will likely lead to the best outcome.
Machine learning clearly has a lot to offer the world. What’s it going to take to realize its full potential?
Typically, people who start with machine learning, and in particular deep learning, try it and it doesn't work. Then they sit down and learn which knobs (also called hyperparameters) to fiddle around with, and after a lengthy trial-and-error process, at some point, it works awesome.
But it's a bit of a black magic. We're trying to go away from this ‘art’ of deep learning, because it means you need a lot of expertise and experience. The more you work with it, the better you get. But it's all unwritten knowledge. Instead, we're trying to get to a principled engineering science where you just say, 'Well, if your data is like this, then you should do this and this and then you will get good performance.' Then we can codify these best practices into an algorithm and everyone can benefit from it automatically. That’s what we’re working towards.
Domain scientists are typically held back by the fact that they're not machine learning experts. They need to figure out which algorithms to use with which hyper-parameter settings. There's gazillions of possibilities, and typically, they don't know what to use and so end up using a default choice that isn't the best for their approach. So, there's a lot of performance potential being lost these days.
But if domain scientists were to use automated methods, like our AutoML tool Auto-sklearn, for doing the machine learning part of their data analysis, then they could come up with much better solutions. With these automatic methods, I think we can really help domain scientists do their job a lot easier. And ultimately, that would democratize machine learning so that anyone can use it to its full potential.