About ten years ago, machine learning technology ushered in a new era in Artificial Intelligence (AI). The neural networks that these algorithms use are often compared to the human brain. But a human child learns what a cat is from three examples; a neural network needs millions of cats.
Bernhard Schölkopf: These methods are very good if you have huge amounts of data and powerful computers at your disposal. But while they start from scratch for every problem, a few cat examples are enough for us humans because we may have already learned how to recognize dogs and how a dog‘s appearance changes under different perspectives and lighting conditions. We haven‘t really figured out how to make that work for computers.
Computer algorithms learn by discovering statistical correlations. But a correlation between two quantities does not establish causality. You are interested in how to detect cause-and-effect relationships in data.
Schölkopf: It‘s impossible with data alone. But it is possible if you can intervene in a system – then you can wiggle one variable, so to speak, and see how the other one changes. If this is not possible, it may be sufficient that objects or conditions change by themselves. My team has applied this to the problem of finding exoplanets and we have actually discovered a number of those planets in the Kepler telescope data. Coincidentally, the first exoplanet we found was in the habitable zone around a star, and water vapor was found spectroscopically in its atmosphere.
For the past year, ChatGPT has been making headlines. You can almost talk to the system like to a real human. Is this technology the future of AI?
Schölkopf: On the one hand, it‘s absolutely fascinating how well these systems work. These large language models have absorbed the cultural achievements of humanity that have been written down as training data and are a kind of distillation of this knowledge. Because it‘s almost like chatting with a human, we are tempted to attribute properties to those systems that they do not have. There is a danger of thinking that the problem of intelligence is solved and we understand how it works.
Training these systems costs billions, and only a few large companies, all in the USA, can afford it. Does Europe stand any chance at all against these AI superpowers?
Schölkopf: I see three possibilities. One option is to start from scratch – but a university or even a Max Planck Institute doesn‘t have the computing power required for that. You could also try to improve an existing model by fine-tuning it for certain problems. Or you simply use the interface to OpenAI or Google. But we should not just chase after the Americans, we should make sure that innovation happens in Europe as well.
That was probably one idea behind Cyber Valley which you co-founded in Tübingen.
Schölkopf: In the past, the best graduates of our universities usually aimed for a career in science. Today, many want to go to one of the top industrial labs. If you really want to be attractive as a region, you need a mix of academic research, industrial labs, and startups. That‘s what Cyber Valley is supposed to provide.
But there is also the European level – and that is where the ELLIS Institute in Tübingen comes in which was launched this summer under your leadership.
Schölkopf: There are other great research locations in Europe, for example around Cambridge and at ETH Zurich. We want to connect those locations in a meaningful way, for example by supporting PhD students with supervisors in two different countries. Such an ecosystem would be a bonus which you would not get in the same way in America. And our students are certainly just as smart as American students.
The interview was conducted by Christoph Droesser