The Highlights of our Thought Leader Interview with Dr. Rasmus Rothe
on Leadership and the Rise of Intelligent Machines
What is the state of machine learning today? Is it just a hype or here to stay? And how should we as a society deal with the tremendous developments that have been made in the previous years? The following interview was conducted by Christian Spier and Dr. Laura Bechthold as part of the Future of Leadership Initative’s (FLI)’s project titled “Yes, Mr Robot! Leadership and the Rise of Intelligent Machine”.
- Machine learning works very well in constrained environments. However, building robust models to cope with edge cases remains the big bottleneck when it comes to deploying AI models to the real world.
- Since the value-add of AI is the highest for complex human decision making, it is paramount for AI startups and developers to closely interact with regulators on setting the right frameworks.
- Machines are very good at solving very specific tasks. However, we are far away from reaching “general artificial intelligence” any time soon.
- Everyone needs to have a basic understanding of AI since it will become part of our daily lives.
- Startups are the perfect bridge between academia and large corporations when it comes to transitioning research into practice.
ABOUT DR. RASMUS ROTHE
WHAT ARE THE MOST DISRUPTIVE DEVELOPMENTS WHEN IT COMES TO THE RISE OF MACHINE INTELLIGENCE?
I think there are a few. Historically, it was already shown in the ’80s that neural networks generally work, and Yann LeCun did some great work on recognizing handwritten digits for instance. But then, we experienced an “AI Winter” in the ’90s and 2000s where those methods became out of fashion. The big breakthrough, however, really started with the Krizhevsky paper in 2011 where, for the first time, a deep learning-based algorithm was by far the best-performing method at the ImageNet competition. This milestone has really proven the potential of deep learning solutions and I think that three reasons have fueled this development. Firstly, data availability – access to data to train networks not a few thousand but on a few million images. Secondly, computational power – the power to process large networks and learn complex relationships. And thirdly, breakthroughs of the scientific community in the late 2000s that allowed us to train large networks on large datasets. I think that this was really the beginning of the recent AI hype that has stayed for the last decade.
[Authors note: ImageNet is a large visual database designed for use in visual object recognition software research. It is also the host of the “ImageNet Large Scale Visual Recognition Challenge” where Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky from the University of Toronto submitted a deep convolutional neural network architecture called AlexNet which achieved a 15.3% error rate in the 2010 contest, by far beating the second place that achieved a 26.2% error rate. Read the full story on QZ and find the Krizhevsky paper here]
WHAT DO YOU THINK ARE THE MOST IMPORTANT AREAS OF BUSINESS AND SOCIETY THAT ARE AFFECTED BY THESE DEVELOPMENTS?
I think that AI will basically impact every single industry, in particular healthcare and mobility. For one thing, healthcare is very close to our hearts, a large industry globally, and also one where you have a lot of data. However, the way a doctor operates nowadays is often not very data-driven and provides space for advanced machine learning, for example in the context of radiology or pathology. Generally speaking, machine learning provides the opportunity to improve patient treatment, reduce operations cost, and ultimately provide access to those who don’t have access to healthcare services right now.The second big area is mobility. Whether it is autonomous cars, drones, or logistics more broadly, autonomous mobility and additional services such as food delivery will have a huge impact on our society. In 2005, when the DARPA challenge was completed successfully by Sebastian Thrun and his team, people thought that it would just take two or three more years for autonomous driving would be widespread. After almost fifteen years and multiple billion dollars invested, we are still not there yet and it will still take a few more years. But once we are at a stage where autonomous mobility will take hold of a certain region, we’ll witness a huge impact on the number of cars on the street, access to mobility, and the cost for ridesharing services.
WHAT IS THE BIGGEST CHALLENGE TODAY?
If we look back to the 2011 competition, the difference between academia and real-world applications comes to mind. In academia, you look at average performance metrics and try to improve them. But for a real-world system, you need to make sure that it works all the time or at least know which situations it can’t handle so that the driver can take over for instance. If you want to deploy a machine learning algorithm in the real world, you need to make sure it can deal with all kinds of rare situations and teaching that to your neural network is quite hard.
ABOUT THE PROJECT
IMAGINE 10 YEARS FROM NOW: WHAT IMAGES AND MOST SIGNIFICANT USE CASES COME TO YOUR MIND WHERE OUR SOCIETY COULD BENEFIT FROM ARTIFICIAL INTELLIGENCE? WHAT COULD BE BETTER THAN TODAY?
Any process where a human looks at data and makes a decision, is ultimately something that can be automated through machine intelligence, thereby making smarter decisions, decreasing costs, and increasing accessibility. Beyond the use cases we’ve already talked about, applications in law and finance for instance are helping to take out human biases. Apart from that, AI is already helping to eliminate repetitive human tasks, similar to the industrial revolution but even to a greater extent.
WHAT IS YOUR IMPACT IN SHAPING THIS POSITIVE FUTURE?
At Merantix, we have been incubating companies for almost four years now. Our automotive company SiaSearch builds an intelligent data processing framework that dramatically accelerates development time for autonomous driving, while our healthcare company Vara has developed a medical imaging platform that significantly augments radiologists’ detection rate and already obtained regulatory approval. Our third company Merantix Labs is a more exploratory arm that runs projects with industry partners. As we are speaking, there are plenty of new companies underway. Our next company is a biotech company but we are also looking into cybersecurity, business intelligence and manufacturing.
Another way in which we hope to have an impact is the Association for German AI Startups (“KI Bundesverband e.V.”) which now comprises more than 200 startups and was built to establish ties with policymakers and make sure that the regulatory and funding environment is active so that we can build more successful AI startups in Germany.
LET US MOVE TO MORE CRITICAL ASPECTS OF MACHINE INTELLIGENCE. LET US THINK AHEAD AGAIN. IN WHICH AREAS OF SOCIETY COULD MACHINE INTELLIGENCE HAVE A NEGATIVE IMPACT?
When imagining negative future scenarios regarding AI, many people think of science fiction scenarios with machines being as smart as humans – what we call “general artificial intelligence”. However, we are far away from general AI and are much better at the more specialized “narrow artificial intelligence” where we use an algorithm to solve one specific problem, throw more data at the algorithm than a human could ever consume, and it will be better than a human at this specific task. Having said that, machines have become very good at generating realistic data. One method behind that is the General Adversarial Network (GAN) which for instance can generate realistic images of people that don’t really exist. OpenAI published a paper where they generated coherent text with minimal prompts. Those developments have fueled the discussion around fake news and fake content which might certainly pose a threat to society.
WHAT IS YOUR OPINION ON THE ISSUES ON BIASES AND EXPLAINABILITY OF AI ALGORITHMS?
Interestingly, the topic of biases is a very old one, yet currently very transparent when it comes to machines making the decisions. If you think about a bank in a rural area, the decision of whether someone gets a loan is also dependent on many factors that make the clerk look biased. If we take it to an extreme, a completely bias-free model would have to be 100% random in choosing the ones who get a loan otherwise we could always argue in one way or the other for some kind of bias. And it behaves similarly with machines unless there we can control the extent of bias. However, the problem is very complex which makes it especially hard for regulators to act upon.
In terms of explainability, the nature of the problems that you are trying to solve with deep learning is very complex. And you cannot expect to solve a very complex cognitive task with a model that is easily explainable. If that were the case, you could also build a simple model that solves the problem. Therefore, I think that people need to get used to the fact that models are not always explainable and trust in rigorous statistical testing prior to deployment of a model.
WHAT ARE VITAL QUESTIONS BUSINESS LEADERS HAVE TO ASK THEMSELVES, THEIR TEAMS, AND THEIR ORGANISATIONS FOR THE FUTURE
Since every leader in every industry will be affected by AI in the future, it is vital that they get a general understanding of the technology and its use cases. Even a three-day crash course would be very beneficial for them to understand the basic concepts behind AI.
WHAT WOULD YOU TEACH IN SUCH A CRASH COURSE ON AI?
Understanding what’s behind the buzzwords: supervised/unsupervised machine learning, data quality (labelled/unlabeled, noisy/clean, standardized, etc.), what does really work today (image recognition, language understanding, numerical methods, etc.), challenges of putting a machine learning model into production, explainability, statistical testing.
WHERE DO YOU SEE DIFFERENT STAKEHOLDERS COME TOGETHER TO SOLVE SOCIETAL CHALLENGES BROUGHT ABOUT BY AI?
I think that startups are the perfect bridge between academia and large corporations when it comes to transitioning research into practice. The US and Asia are more advanced in this regard, but we need to foster our startup ecosystem, closely interact with the regulators, and make sure that the broader population understands the true potentials and dangers of artificial intelligence.
WHAT IS YOUR RECOMMENDATION FOR YOUNG TALENTS HOW TO PREPARE THEMSELVES FOR A FUTURE WITH AI?
Regardless of your study background, you need to have a basic understanding of AI since it will be part of your daily lives. Generally speaking, it is a great time to be in this space since demand is huge and expected to rise so that you can also take part in shaping the industry yourself.