The A.Ideas series presents opinion pieces based on the discussions from the OECD Conference "AI: Intelligent Machines, Smart Policies". This article is based on a position Paper from Philipp Slusallek, Scientific Director at the German Research Center for Artificial Intelligence (DFKI), Saarland University
Future Artificial Intelligence (AI) systems need to interact safely and reliably within a highly complex real world. This includes applications like autonomous driving, robots interacting with their environment, but also decision making under complex constraints in areas such as social sciences, finance, politics, and many others.
To achieve this, AI systems need to simultaneously understand the real world (by learning models or the “rules of the game”), as well as find the best strategies to act given these models of reality. Machine learning is the best method for such systems to adapt to and learn about complex interactions with the real world. This typically requires large data inputs for training and does not offer methods for the formal validation of results, both of which cause severe challenges in many areas. For example, how can we guarantee an autonomous car will exercise the correct reaction when faced with the risk of hitting a child running across the street, when we (obviously) cannot train and validate their performance in the real world? We need to be sure that our AI systems are trained to do the right thing under all circumstances. And not just for autonomous driving but for many, if not most, other areas of our daily lives.
Fortunately, both problems can be addressed by developing a “digital reality”: a simulated environment that faithfully replicates all relevant features of the real world in a computer. This digital reality can then be used for future AI system behaviour training, benchmarking and for performance validation (e.g. a rigorous “driver's license test” for autonomous vehicles). Even better, such systems can be used to explore and reason about the decision-making process and its potential outcomes – thus it may be able to help inform about possible strategies and desirable policies.
This approach is similar to other recent success stories from AI, where for example AlphaGo Zero has been able to learn from scratch to superhumanly play Go within just a few days of training. It did so mainly through simulation, simply by playing Go against itself. It started with random movements, but it learned which moves are more helpful in certain situations and those that are not by observing their medium- and long-term outcomes (via Deep Reinforcement Learning). Moreover, it even went a step further and discovered new game strategies that were not known to human Go players before.
Applying this approach in the real world could be very helpful. But manually creating a digital reality for these purposes would not scale and would not be able to include the subtle details that are important for many tasks. In contrast to Go, the “rules of the game” in reality are much more complex and are not readily provided – we must first discover them. This essentially means we need to create models of the real world e.g. the motion of pedestrians, when pedestrians decide to cross a street on a given traffic, how traffic and environment impact one another, how traffic behaviours change with the weather etc.
Digital reality training loop
We propose to use machine learning by observing the real world to automatically and continuously learn and update a digital reality. The digital reality training loop from Figure 1 shows the general approach for autonomous driving. We examine partial models, one for each aspect (e.g. pedestrian motion or traffic environments) from data that is readily and ubiquitously available in the real world. We can then combine these different partial models to describe entire scenarios, including critical ones for which we otherwise would have no data such as a child running in front of a car. However, for proper training via machine learning, we need not only a single instance of this scenario but a representative set of all possible variants. Given that we have models that describe the different aspects of reality, we can automatically create any variation (e.g. large or small kids, different clothing, coming from left or right, fast or slow, at day or night). We can then simulate any of these variants and obtain the required synthetic sensor data for cameras, radar, lidar, etc. (essentially all the data a real car would observe in reality in those situations).
Just as with learning to play Go, we can then explore what a car could do in each situation. In simulations, we can determine to what degree each such “move” is helpful or not in the medium and long term, and thus learn and improve strategies for acting; that is, to drive safely and reliably.
Continuous validation and adaptation
One key difference to playing Go, though, is that our model of the real world might not be perfect. As a result, we need to make sure that what we learn in a digital reality reflects the real world as accurately as possible. This can be achieved by training our system with both synthetic and available real-world data simultaneously.
Furthermore, we can detect possible mismatches between our models and reality. By updating our models or our learning processes whenever these mismatches are detected, we can continuously adapt and improve our understanding of the world. Obviously, this entire approach is not just limited to training autonomous driving but can be applied to nearly any application for which we can learn and simulate suitable models of the real world.
Do we need a CERN for AI?
Together, AI and digital realities allow for the continuous and simultaneous improvement of our understanding of both the real world and of how to make good decisions based on this understanding. Similar to the way computers can now learn to play Go through simulations, these methods will allow us to build AI systems that can interact with the real world on our behalf, well beyond what we can currently imagine.
Progress in AI must benefit everyone, creating an “AI for Humans”
This will be a real game changer: it is highly challenging scientifically; will enable completely new research fields across all disciplines; and requires scientists from many different areas to work together, as well as technologies and datasets to be integrated into a common platform. As a result, completely new business models and new types of companies could develop and society, in general, could benefit from better solutions for complex local and global problems. The integrated platform should be as open and flexible as possible to promote research and immediate experimentation. But it also needs to facilitate the bidirectional transfer of results and datasets to and from industry to encourage commercial applications and spinoffs. These capabilities should not be locked behind closed doors but developed with full transparency, including open data where possible. It is essential for the public to know what progress is being made and how the results can and should be used. Progress in AI must benefit everyone, creating an “AI for Humans”.
It is obvious that building such an AI platform is a challenging task that no individual organisation can tackle alone. Creating this platform would involve the development of a methodology, plus learning strategies and models that reflect the real world. These models would then need to be combined to simulate more complex scenarios, covering real life situations and adapting as they change, and collecting the data needed to support such activities.
But first and foremost, such a platform needs to facilitate the communication and exchange of concepts and information in the community. For autonomous driving alone this includes aspects such as: suitable models of the environment; how each of the objects responds to light, radar and other measurements; how objects move and behave; lighting and weather conditions, and much more. For robotics we also need to include the physics of interacting with the world. For social sciences, we need to model the many and complex characteristics of human interactions. For policy decisions, including the economic and financial sector, many of these aspects need to be looked at together.
Fortunately many researchers across disciplines, as well as companies, are already addressing many of the individual challenges of modeling our world at various levels of detail and accuracy. However, today most of these activities are largely done in isolation, siloed by groups creating their own methods, infrastructure and datasets. There is currently no systematic approach nor a common platform for bringing these different pieces together.
This is where AI can learn from other disciplines. For example, the Human Genome project has brought researchers from around the globe together to jointly and systematically study human genomics. As a result, we have made huge progress in just a few years that would have been impossible without these co-ordinated efforts.
Similarly, physics has a long tradition of systematically organising its key research initiatives around large, strategic projects. A prime example here is the study of fundamental particle physics at CERN: the particle accelerators, detectors and computing infrastructure that have been brought together in an effort to support the research of thousands of groups worldwide are remarkable.
As a result of these considerations, we suggest the establishment of a CERN for AI: a collaborative, scientific effort to accelerate and consolidate the development and uptake of AI for the benefit of all humans and our environment. While the details will obviously be quite different, we use CERN as a model to communicate the spirit of our research platform.
Similarly to CERN, the mission of this AI platform would be to continuously improve our understanding of the world around us and use this information to explore and evaluate better ways to act and interact in this world. This would be a highly interdisciplinary endeavor bringing together scientists from many areas in close collaboration and interaction with industry, politics and the public. A key element would be a common reference platform for AI and digital reality, where we can bring together researchers to jointly advance the many strands of research in this area by discussing common approaches, methods, algorithms, data and their applications. This platform should be open, flexible and facilitate the transfer and exchange of knowledge to accommodate a wide range of research, including industry efforts. We believe that similarly to CERN, a high degree of self-organisation and pragmatism around a central mission and infrastructure will be instrumental in adapting quickly to future challenges.
The German Research Center for Artificial Intelligence (DFKI) is prepared to help drive this vision.
GOT A FEW MORE MINUTES?
Artificial intelligence and the right to informational self-determination, by Mathias Cellarius
ABOUT THE AUTHOR
Philipp Slusallek is Scientific Director at the German Research Center for Artificial Intelligence (DFKI), where he heads the research area on Agents and Simulated Reality. At Saarland University he has been a professor for Computer Graphics since 1999, a principal investigator at the German Excellence-Cluster on “Multimodal Computing and Interaction” since 2007, and Director for Research at the Intel Visual Computing Institute since 2009. Before coming to Saarland University, he was a Visiting Assistant Professor at Stanford University. He originally studied physics in Frankfurt and Tübingen (Diploma/M.Sc.) and got his PhD in Computer Science from Erlangen University. His research covers a wide range of topics including artificial intelligence and multi-agent systems, digital reality, real-time realistic graphics, motion synthesis, novel programming models, high-performance computing, computational sciences, 3D-Internet technology, and others.