Several systemic research fields, which pose central questions on the understanding of complex systems, from recognition, to learning, to adaptation, are investigated within the Max Planck ETH Center for Learning Systems:

• Processes of (self-)organization, (machine) learning and artificial intelligence of complex systems.
• Technologies to probe intelligent biological systems and their ability to adapt to varying external dynamics, including the nervous system and new computational, mathematical and robotic models of such systems.
• Advanced technologies – from micro to macro – to design, fabricate, assemble, control, and interrogate complex synthetic, bio-inspired or bio-hybrid systems.
• Robust model-based control of intelligent behavior.
• Stimulus-response, as well as perception-action-cycles of autonomous systems.
• Learning to perceive in and adapt to complex time-varying environments.

ETH Zurich and the Max Planck Institute for Intelligent Systems have strong groups with an international visibility in all of these areas. The joint research center allows these groups to join forces with the common goal to investigate, understand and master the challenges of intelligence, learning and control in the zettabyte age.

G1 Machine Learning and Empirical Inference of Complex Systems
Machine learning and empirical inference address scientific questions of how statistical models should be designed, estimated, and validated based on massive data arising on multiple scales. The current trends in learning systems cope with high dimensionality using sparse models or non-parametric Bayesian approaches. Many of the models used in practice for data mining applications, e.g., document analysis, exploration of biomedical and systems biological data, as well as online monitoring of sensor data and security information, are characterized by high complexity and few data points relative to the number of degrees of freedom. Novel ideas in statistical learning theory are required to cope with the resulting curse of dimensionality and the high measurement uncertainty. The success of ensemble methods such as boosting, bagging, and the dynamic model adaptation in non-parametric Bayesian estimation highlights the impact of a new class of data driven adaptive models with a dynamic complexity control. Computational questions, which played a secondary role in classical statistics, substantially gained in importance for current research on very large data sets with hundred millions of records and more. Mining of large data sets demands an efficient information system infrastructure as well as the development of robust statistical models. Efficient algorithms with strong performance guarantees will be of crucial importance to enable these applications.

G2 Perception-Action-Cycle for Autonomous Systems
A common feature shared by most autonomous systems is the concept of a perception-action loop. Biological systems interact with the world by perceiving relevant aspects of the world and their own state, processing these aspects, continuously deciding what actions to take based upon this information and adapting to the environment by updating their priors. While the computations involved in the processing of perceptual data and generation of actions - and in particular the closed-loop properties of perception-action systems - are far from understood in biological systems, it is clear that they involve aspects of learning and adaptation as well as processes of inference in the face of uncertain and hidden information derived from very high dimensional multi-modal data streams. Moreover, perception, action, and learning are tightly linked into a functional system, i.e., it is hardly conceivable that the building blocks were developed independently of each other. Thus, studying complete perception-action-learning loops, not in a component-wise isolation but rather as an integrated system, emphasizing tools from empirical inference, will most likely form a new focus of autonomous systems research. This focus will also require a novel kind of interdisciplinary researchers, i.e., people who can work on action, perception, and learning simultaneously and understand the interdependencies of the different fields, rather than having pure action specialists, perception specialists, and machine learning specialists.

G3 Robust Model-Based Control for Intelligent Behavior
Using models to achieve high-performance control systems is a typical approach in control engineering. Linear control is particularly well developed, but also nonlinear control has received significant progress in the past 20-30 years. Models can be built over any relevant control aspect of the control system and its environment. Typical models include dynamics models, kinematics models, friction models, predictive models of dynamic events in the environment, etc. A good model provides more control accuracy, faster reaction time, and reduced reliance on strong error correcting mechanisms, typically implemented as negative feedback controllers. For instance, in assistive robotics it is desirable that robots in interaction with humans are compliant, i.e., they can gently give-in when unexpectedly impacting with the environment (which might be a child). Such compliance requires high-quality model-based controllers, as only these controllers can minimize the stiffness that would otherwise be created by a negative feedback controller with high gains. Methods to learn and adapt this compliance based on the external dynamics must also be further developed. Exploring how robust model-based control for intelligent autonomous behavior can be created remains an open and most relevant research topic. Various venues can be explored, leveraging, for example, robust, stochastic, and optimization-based control design. Building models with tools from empirical inference is a promising route, but needs to be reconciled with the desires of control engineers to have provable performance bounds.

G4 Robust Perception in Complex Environments
While perception, action and learning are the foundation of autonomous behavior, biological systems have vastly different perceptual needs and strategies. Nature has evolved a staggering array of sensing systems including many type of “eyes”, as well as auditory, haptic and olfactory systems. While a single unified “theory of perception” across systems and scales is unlikely, the study of perception at different scales is needed to understand autonomous behavior. What are the requirements on perceptual systems at different scales? What sources of uncertainty exist? What strategies are used to deal with complexity and uncertainty? How is perception integrated with behavior? The novel approach of this Center is to study perception across scales, from human-like vision to the cellular level, in a coordinated fashion to explicate the commonalities and differences of such systems. At the macro level, complex visual systems in humans, home-service robots, and autonomous vehicles share the luxury of immense computational resources but must act on timescales on the order of 10's or 100's of milliseconds. These perceptual systems are typically characterized by rich signals (e.g. high-resolution color imagery), three dimensions (e.g. stereo depth perception), and tremendous generality, enabling them to adapt to new and vastly different environments. After 30 years of research, computer vision systems have started to have an impact on peoples' lives yet they still lack the ability to learn and generalize at the level we expect from a human toddler or even a family pet. Current systems lack rich representations of the visual world that are learned from visual experience and that can be generalized to new contexts. Clearly new foundations and theories are needed that are strong enough to capture the complex appearance of objects in real scenes and grounded in machine learning.

G5 Design, Fabrication, and Control of Synthetic, Bio-Inspired, and Bio-Hybrid Micro/ Nanoscale Robotic Systems
A major challenge is the design of approaches for building synthetic milli- and micro-devices that can perform specific functions which are in part as successful as cellular ones, e.g., robotic microsystems can be functionalized with biological molecules to either render them biocompatible for human applications or to attribute functionalities that go far beyond those exhibited by synthetic systems, for example the integration of biological nanomotors. These microdevices may lead to amazing new complex materials systems that mimic specific properties of living cells and functional tissues, e.g. self-assembly, self-propelling, and environmental sensing. Moreover, synthetic mimicking approaches in relation to cell biology will not only yield synthetic biomimetic systems, but will also help to develop a deeper understanding of the individual functional units in living cells and the cell itself.

G6 Neurotechnology and Emergent Intelligence in Nervous Systems
Nervous systems will continue to provide a rich source of inspiration and yield deep insights when it comes to the design of information processing algorithms and the engineering complex sensorimotor control loops. Whereas in the past our ability to measure and control brains was limited to electrophysiology, electrical stimulation, and neuropharmacology, in recent years two new techniques have emerged that will propel our understanding of brain functions in health and disease. These techniques are optogenetics and neural circuit reconstruction. The former technique, optogenetics, is mostly about controlling the excitatory or inhibitory state of a single nerve cell or a group of cells with the intermediate of light. The latter technique, circuit reconstructions (using electron microscopy), is about exhaustively reconstructing the networks of excitatory and inhibitory synaptic contacts between cells, to yield a full connection diagram akin to a blueprint of the nervous system (or brain area).