He teaches robots to think and won the second highest award in computer science for his pioneering research in artificial intelligence. Thanks to his physics teacher at Brasschaat College. Breakfast with De Tijd
A colleague’s wedding in Santorini last month was Pieter Abbeel’s first chance in almost three years for a short stay in his homeland and village. “My dad pointed out that it was October 2019,” says Abbeel, who has lived and worked on the U.S. West Coast for over 20 years. The pandemic got in the way, of course, but it was also an opportunity to be hyperproductive. “I recently realized that I have been working every day since the start of the corona,” he says as we sit down at the buffet at the Van der Valk hotel Dennenhof in Brasschaat. For a Monday morning and given its location on the Bredabaan, the hotel has a very extensive breakfast offer.
Breakfast with Time
Brasschaat, kl. 9, Van der Valk hotel Dennenhof
With Pieter Abbeel, we talk about deep learning, injuries at Amazon and Elon Musk’s passion
Time management is crucial if you are Pieter Abbeel and need to decide what to use your brain for. The 44-year-old academic currently combines four jobs, albeit all related to his expertise: artificial intelligence (AI). Abbeel has been a professor at the University of Berkeley, California since 2008, where his research is one of the driving forces behind the impressive advances AI has made over the past decade. He belongs to an international circle of academics who are constantly pushing the boundaries in the field. In his lab, for example, Abbeel and his team managed to teach a robot how to pick up a towel from the floor and fold it. It sounds like a toddler’s task, but for an unprogrammed machine, it’s a huge leap.
In early April, Abbeel received the ACM Award in Computing for Breakthroughs Like This, the second highest award in computer science. The highest, the Turing Prize, is widely regarded as the Nobel Prize for the field. “Abbeel has contributed to major leaps in research, while generously sharing his knowledge to bring robots to exciting new levels of opportunity,” the Association for Computing Machinery said at the award.
Five years ago, it was time to put that research into practice. Together with some ex-students, he founded the start-up Covariant, a company that develops software for automating factories and department stores. In addition to being chairman, Abbeel is also the chief researcher and provides primarily strategic advice. Since its inception, the startup has already raised $ 147 million and expanded to 140 people. ‘The mission is: to make robots as reliable as possible in a context of constant change’, he explains above a well-stocked plate of scrambled eggs, fruit and yoghurt.
“When we started, we were at 95 percent accuracy. It’s now increased to more than 99.5%. It’s a degree of autonomy where a person hardly needs to intervene, and then it’s very attractive commercially. ‘ The machines are not programmed to always perform the same movement as in a car factory, but they understand their environment and purpose so well that they can pick orders independently in e-commerce warehouses. “We can automate very ungrateful jobs, where 70 percent of “People suffer from muscle damage because they have to do the same repetitive movements all the time.”
Job three joined them last year when Abbeel exchanged ideas with an acquaintance from AI circles during a barbecue in his garden. They both experienced that they were inundated with questions from AI start-ups for advice and for investment. This led to the idea of creating AIX, a fund that allows them to get into young growth companies that use AI for a variety of purposes. AIX raised 50 million euros from a few wealthy people in the industry and already has more than 40 start-ups in its portfolio. These include a company that uses image recognition to count white blood cells in blood samples, or one that detects possible health problems in babies based on their crying.
And job four is a corona project that has gone awry. In the beginning of corona lockdowns, Abbeel looked for a way to motivate his students and staff and organized Zoom conversations with other AI leaders. And why not expand it to a wider audience? This is how The Robot Brains was born, a weekly podcast in which Abbeel receives fascinating guests to tell about their careers. Like Andrej Karpathy, head of AI at Tesla, or Geoff Hinton, the godfather of modern AI. ‘Thanks to him, a revolution in deep learning took place in 2012, making almost all important applications possible today. He started in 1971 with computers with neural networks based on how our brain works. For 40 years he was an underdog about who everyone said: Too bad, but it does not work. Until it turned out that he was right all the time. “
For Abbeel, all his activities can be reduced to ‘a great passion’ for artificial intelligence. ‘Thinking separates us as humans from other species. The question that has fascinated me for a long time is: how can we build something that can be just as smart or even smarter than humans? When my career started 20 years ago, it was primarily an academic issue. Companies were hardly involved in it because there were almost no practical uses. It started to change very quickly ten years ago. ‘
Abbeel remembers a lot of Lego (‘first with manual, then without’), puzzles and daily board games with his parents from childhood in Brasschaat. ‘Cluedo, Stratego, Risk: things that required you to think actively. I wanted to learn. In school, especially math and physics because it’s so basic. I have always looked forward to the physics lessons with teacher Nick Merlin here at St. Michael’s College. ‘ He then went to Leuven to study civil engineering, where he eventually took an optional AI course. Then I knew: that’s what I want. But then I had to go to one of the four relevant American top universities’. He earned a master’s degree in computer science from Stanford, California, where he became a Ph.D. at Andrew Ng, the founder of Google Brain and former AI chief at China’s Baidu. Then Abbeel crossed San Francisco Bay to Berkeley, where he became a professional as a 30-year-old.
We teach machines to make connections.
Abbeel’s specialty is what is jargon called ‘reinforcement learning’, a form of deep learning where machines learn things in a way that mimics human learning. Deep learning is the type of AI that works on the basis of artificial neural networks that are inspired by how our brains work. Abbeel explains it in plain language: ‘The neurons in our brains think. They receive input, process it and generate output. There are hundreds of billions of neurons, but each neuron has on average only 100 connections to others. From that network of connections, we make calculations in a process that we still do not fully understand. But when you see a dog, you know it’s a dog because you’ve learned to make those connections. We do the same thing with artificial networks: we teach machines to make connections. And it can be done in different ways. ‘
The latest generation of this is reinforcement learning, where a robot creates connections with a constant feedback cycle and a specific goal in mind. That way, he learns to perform tasks. ‘If a robot wants to pick up a package, it needs to be able to make plans and not just make one decision, and that’s it. Basically, it’s about reasoning about possible futures and then choosing an action plan that generates the most desirable future. ‘
In 2016 and 2017, Abbeel spent a year and a half at OpenAI, the research institute largely funded by Elon Musk, whose goal is to develop AI that benefits all of humanity. Besides all the noise that the man generates because he does not stay out of the news for a day, there are two things that stand out in Musk, says Abbeel: passion and vision. “Because he sets such huge goals and is so demanding, combined with huge resources of course, he attracts certain people and makes them push boundaries. When you start working for Tesla, you know that more is expected of you than Google or Facebook. ‘
Elon Musk makes people push boundaries.
“And he will always understand everything. At OpenAI, he didn’t just stop by to check if everyone was doing the right thing. No, he asked questions like: what are you working on? Why? Why that way? What is the greatest possible success of it? That thoroughness makes a big difference. You can also keep asking him: why, why, why? He will continue to respond. That is not true for most people. Then you eventually bump into: because it’s my job or something. And that’s fine too. But then you do not change the world.
Typical of artificial intelligence is that it progresses in cycles. There have already been so-called AI winters where research did not move forward and so little money was released. Five years ago, there was a lot of media hype – Abbeel was then thoroughly interviewed by the music magazine Rolling Stone – among other things about self-driving cars. But it has settled down a bit. Does he feel cold again on the research side? ‘You need to look at the basic drivers: the amount of computing power and the amount of data. And also the people who work on it. Media attention may come in waves, but these three factors continue to rise very sharply. So I expect a lot of progress in the next ten years. ‘