Mastermind of active machine learning

For Andreas Krause, artificial intelligence and machine learning are two of the most exciting topics of our time. He has helped develop and define a whole host of approaches, ranging from the theoretical and mathematical foundations to practical questions regarding how humans and learning systems interact.

“Humans are very efficient at learning to solve even complex tasks,” Krause explains. He is Professor of Computer Science at ETH Zurich and one of Europe’s leading researchers in the rapidly evolving field of machine learning. “I’m keen to understand how we can enable machines to learn with similar efficiency,” he says.

Krause first came across the topic of machine learning during his Master’s degree, which he completed at the Technical University of Munich in 2004. Thereafter, to gain a deeper understanding of the principles, he went to Carnegie Mellon University in Pittsburgh, where he wrote his doctoral thesis on issues of optimal information gathering. After completing his doctorate in 2008, he became an assistant professor at Caltech in Pasadena in 2009, before joining ETH Zurich in 2011. Since then, he has also taken on leadership roles at the Swiss Data Science Center, the ETH AI Center and the European Laboratory for Learning and Intelligent Systems, ELLIS.

To honour his achievements, he was awarded the 2021 Rössler Prize on Wednesday at the ETH Foundation’s thanksgiving event. “Andreas Krause is a stellar researcher and a dedicated lecturer, and for someone at an early stage of his academic career, he has already earned a notable number of merits in one of the most impactful technologies of the 21st century,” said ETH President Joël Mesot in his speech.

The fascination of learning agents

Krause is particularly fascinated by questions of optimal information gathering that require efficient, active forms of learning such as reinforcement learning. The key factor here is dealing with uncertainty when not all the information is yet available or when there are a multitude of alternative solutions.

In conventional “passive” machine learning, a learning algorithm is trained using large data sets to deduce certain patterns from data annotated by experts – enabling, for example, a classification rule to be identified that recognizes whether photos show pedestrians or traffic signs. “Active” learning methods on the other hand decide for themselves what data they require to perform the task in hand effectively. For instance, they can go through an image data set, select any images that will help learning progress and have them annotated by experts, which can save time and money and reduce errors. Other active methods propose experiments whose outcomes promise valuable information.

Such computer programs are also known as learning agents. “I'm fascinated by active learning methods, where a learning agent decides for itself which data will be useful in helping it to make good decisions,” Krause says.

Robotics is one field in which such issues arise. Krause illustrates this using a drone that learns to solve certain tasks actively by means of experimentation. The difficulty here is that at the beginning, it isn’t possible to say exactly what might cause the drone to crash and what won’t. At first, the drone must behave cautiously; then, the more data it obtains and analyses, the better it will be able to perform without putting itself or others in danger.

Dilemma between old and new data

However, for a learning agent to gather information actively is no trivial process. To optimise its fulfilment of a task, it must find the right mix of existing data and new, self-acquired data.

This trade-off is what researchers refer to as the “exploration-exploitation dilemma”: if the learning agent actively decides which experiments to conduct in order to obtain additional data, then its decisions also influence which data it does and does not have available as it learns.

As one of his seminal achievements, Krause developed the first mathematical learning method for which, based on certain assumptions, it is possible to prove that it will solve the exploration-exploitation dilemma effectively, even in complex applications. Mathematically speaking, this is a variation of Bayesian optimisation, which also underlies the example of the drone learning while avoiding to crash, and which provides certain formal safety guarantees under specific conditions.

Pioneering and passionate researcher and teacher

Krause’s research is indeed very mathematical. For example, active learning methods can utilize very specific “submodular” functions to acquire useful data as efficiently as possible. Today, Krause is viewed as a pioneer who brought submodular optimisation to machine learning. The findings from a very influential publication by Krause from his time in the US even led to practical applications in water distribution networks, addressing the question of where to best place the sensors in order to ensure optimum measurement of water quality.

Krause is not only a sharp thinker when it comes to the mathematical underpinnings of machine learning; he is also someone who reflects on the impact these technologies might have on business and society. For him, it is essential to ensure that the algorithms or calculation rules underlying the learning procedures are reliable, explainable and traceable, and that the results, decisions and recommendations are fair and trustworthy for whoever they affect.

And Krause relays this conviction in his lectures. Alongside the fundamentals of mathematics and computer science, he is passionate about teaching future AI experts a sense of responsibility regarding the use of the technologies. The Golden Owl Award he received from ETH students in 2012 for his teaching and the fact that over a thousand students attend his “Introduction to machine learning” lecture bear testament to his commitment. He was also instrumental in designing the data science Master’s programme and the DAS in data science at ETH; in the ETH AI Center, he ensures that more entrepreneurial aspects are taught so that spin-offs can increasingly put the acquired AI skills into wider practice.