What will the next 50 years bring for EPFL computer science?
EPFL has established itself as a leader in traditional core fields of computer science. School of Computer and Communication Sciences (IC) Dean Jim Larus looks ahead to the challenges he’d like to see tackled, including more interdisciplinary research on scientific and social dilemmas.
New systems of computing that go beyond silicon are going to be needed.
© Alain Herzog
Dean Jim Larus.
Since taking up the Deanship in 2013, Larus has placed a strong emphasis on hiring, especially of junior faculty. Despite having finally reached a “critical mass,” Larus wants to keep growing the department, especially with young researchers specializing in diverse fields. “We need to grow outward, and to broaden beyond traditional avenues of research,” he says, citing recent expansions into data science and machine learning.
Larus strongly supports interdisciplinary study, especially to address social and ethical questions associated with emerging technologies like machine learning and artificial intelligence (AI).
“We’ve seen the dark side of many things people thought were good, like social networks. People worry about AI and ’killer robots’, but there are also questions about jobs: if we have self-driving cars, will we need taxi drivers anymore? A lot of these problems are multidimensional, and you need to talk to experts in other areas, like ethics and the humanities.”
The quantum revolution
Quantum computers, Larus says, are on the brink of becoming a reality; the world needs to be ready for them, and this is another area of research that he would like to see more of at EPFL. Rather than storing information in binary states represented by 0 or 1, quantum computers store information in the ’superposition’ of the 0 and 1 states simultaneously, and consequently can perform many computations simultaneously. Because quantum computers could potentially process unprecedented quantities of data compared to today’s computers, data management systems, algorithms, programming languages, and infrastructure all need to be adapted to support the quantum systems of tomorrow.
The fact that Moore’s Law is coming to an end represents another paradigm shift for computer science. The computing principle states that the number of transistors on computer chips – and overall computing power – doubles every two years, and has held true since the 1960s. But computing power and transistor size are finally reaching their limits, Larus says: “We haven’t been able to run computers faster since 2005; the heat is just impossible to manage.”
In the short term, building more specialized computer processors is one solution. But for the future, Larus believes that entirely new systems of computing that go beyond silicon are going to be needed. “I’m particularly interested in using biological processes for computing. If you look at our brains, they’re amazing computers: they run on 25 watts of power, they’re very adaptable, and they can learn things in a matter of seconds.”
In addition to searching for parallels between the human brain and computers, inspiration can also be found in genetics. For example, researchers are starting to look at compact, stable DNA molecules as the inspiration for long-term data storage.
“I want to hire people who think outside the box. This is what a university should be doing: taking these wild ideas that are plausible, and making them real,” Larus says.
Celia Luterbacher, science writer IC
Security and privacy, from e-voting to genealogy
One thing IC is very well known for is its work in security, privacy and digital trust. “So much of this data is being collected, and we’re starting to see people using it in ways in which it wasn’t intended,” says IC Dean Jim Larus. “For example, in the US, a lot of genetic data is on genealogical websites, and police have started using these sites to find criminals through their relatives.”
The Laboratory for Data Security (LDS) has long been working on wireless security, interpersonal privacy, and medical data. Last year, the LDS released the first operational system aimed at protecting sensitive patient data, including genetic information, allowing it to be used for medical research without loss of privacy.
The Decentralized and Distributed Systems Lab (DEDIS) is also extremely innovative and creative in its work on blockchain. While cryptocurrencies like Bitcoin are perhaps the most well-known application of blockchain technology, the DEDIS lab is also developing security solutions that can be derived from decentralized systems. These include e-voting, and secure data science and data-sharing platforms.
Computer vision for state-of-the-art modelling
In the Computer Vision Lab (CVLab), researchers aim to emulate the brain’s incredible ability to make sense of visual data, for example by recovering 3D motion from video sequences. Computer vision technologies developed in the CVLab have practical applications in many fields, from medical image processing and flying object detection for unmanned aerial vehicles, to augmented reality and the reconstruction of human motion for sports analysis and training.
The CVLab is very active in the transfer of these technologies to industry, for example via collaborations with companies like Flarm. They also have a start-up, Neural Concept, which works on a totally different problem: computational fluid dynamics. Traditionally, a simulation would be used to model the flow of air over an object – for example, to create an aerodynamic shape like a car or airplane – which is very expensive. But it turns out that machine learning (ML) is very good at solving these kinds of problems. Neural Concept has developed an ML-based software application that was used to design an ultra-aerodynamic bike, which this year broke a global speed record.
Machine learning and data science… everywhere
One thing the Center for Digital Trust (C4DT) has made very clear is that machine learning and data science are extremely important already across EPFL; not just in computer science, but also in engineering, life sciences, and architecture.
Machine Learning and Optimization Lab head Martin Jaggi co-teaches Machine Learning, where students are encouraged to collaborate with any research lab on campus on real problems. Last year, this resulted in over 50 interdisciplinary projects on diverse subjects including solar energy, psychology, air quality, classical music and Turkish politics.
Machine learning algorithms are also central to the emerging discipline of digital humanities. Recent work in the Digital Humanities Institute has seen these techniques applied to classical music, experimental museology, and cultural and historical preservation.
The Machine Learning and Optimization Lab (MLO) at EPFL has recently discovered that an unsupervised learning algorithm can be applied to the design of autonomous vehicles, improving safety on road ways in Switzerland. The research, published in the journal Intelligent Transportation Systems, was done in collaboration with the European Transport Safety Council, and it is the first effort to apply machine learning algorithms to autonomous vehicles in Europe.
The paragraph above may sound like an EPFL press release, but it is actually “fake news.” It was written using Transformer, a type of neural network architecture that can be used to automatically generate or translate texts. MLO researchers are working with Transformer models to better understand how they work, and to find different ways to “train” them.