The faculty of the International Max Planck Research School for Intelligent Systems offer numerous courses that may appeal to our Scholars and Associated Students. With few exceptions, the courses listed here take place at the Univeristy of Tuebingen, the University of Stuttgart, or one of the two sites of the Max Planck Institute for Intelligent Systems. These courses typically require in-person attendance.
Click the course name to be taken to the more detailed course listing below:
Advanced ANN topics. First, revisiting backprogation and backpropagation through time; then: Advanced Recurrent Neural Networks (LSTM); Deep Learning; Convolution; Reservoir Computing; Dynamic NNs; Hierarchical Vision Architectures; Restricted Boltzmann Machines; Predictive Encoding & Free Energy; Gain Fields and Switching Networks; Generative Networks.
Programming and design of intelligent, realistic, interesting, behaving avatars, objects, and tools in virtual realities.
Multi-agent systems are systems composed of multiple interacting dynamic units. These units can be used to perform team objectives with applications ranging from formation flying to distributed computation. Challenges associated with these systems are their analysis and synthesis, arising due to their decoupled, distributed, large-scale nature, and due to limited inter-agent sensing/communication capabilities. This course provides an introduction to these systems via tools from graph theory and dynamic systems theory. The course will also cover real-world applications by presenting recent results obtained in the distributed formation control of real multi-robot systems.
Die Studierenden
Based on our knowledge about how animals and humans plan their behavior, make behavioral decisions, control their behavior, and progressively optimize and adapt it, behavioral decision making, control, optimization, and adaptation algorithms are introduced. In particular, the lecture introduces spatial representations for behavioral control, forward-inverse control models, including the learning of such representations and models. Also the encoding and the learning of motor control primitives and motor complexes is considered. Last but not least, self-motivated artificial systems are considered that strive to maintain internal homeostasis and to maximize information gain.
Cognitive models covering learning, action and perception are presented and discussed, including descriptive, qualitative, quantitative and neural models. In addition, parameter optimization as well as techniques to compare models and to interpret and evaluate model parameters are introduced. All techniques are shown in the context of concrete models of cognitive processes.
In our daily live, we are able to perform complex movements and to adapt our movement behavior continuously to changing environments. For doing so, our motor control system performs continuously various control and adaptation processes on different controls levels. In order to be able to understand and to model these control and adaptation processes, basic knowledge on the interaction between neural control and biomechanics is necessary. Goals of this lecture are on the one hand to get an understanding of motor behavior as an interaction of neural control and biomechanics and (changing) environments and on the other hand to learn methods for the development of models of motor control and motor learning processes.
Although the digital photography industry is expanding rapidly, most digital cameras still look and feel like film cameras, and they offer roughly the same set of features and controls. However, as sensors and in-camera processing systems improve, these cameras will begin to offer capabilities that film cameras never had. Among these will be the ability to refocus photographs after they are taken, or to combine views taken with different camera settings, aim, or placement. Equally exciting are new technologies for creating efficient, controllable illumination. Future "flashbulbs" may be pulsed LEDs or video projectors, with the ability to selectively illuminate objects, recolor the scene, or extract shape information. These developments force us to relax our notion of what constitutes "a photograph." They also blur the distinction between photography and scene modeling. These changes will lead to new photographic techniques, new scientific tools, and possibly new art forms. In this course we will survey the converging technologies of digital photography, computational imaging, and image-based rendering, and we will explore the new imaging modalities that they enable.
Non-smooth dynamical systems are in the meanwhile well accepted as a mathematical tool which provides an adequate description for many applications originating from such areas, as electronics (any kinds of switching circuits) and mechanics (impacting and friction dominated systems) as well as social sciences (systems including decision making) and economics (business cycles models). The goal of this lecture course is to present a broad overview of possible effects both from the theoretical and from the applied perspectives. We consider the classes of attractors and other invariant sets which appear in such systems, as well as their possible bifurcations (border collision bifurcations, degenerate bifurcations, homoclinic bifurcations leading to transformations of chaotic attractors).
(Vorlesungsverzeichnis Nr. 05007) für Studierende der folgenden Fachrichtungen: Chemie-Bachelor, Chemie-Höheres Lehramt, Lebensmittelchemie, Chemie-Bachelor of Arts, Materialwissenschaft, Mathematik und Technikpädagogik
Reglerentwurf für lineare zeitinvariante Systeme im Zeit- und Frequenz-bereich. Insbesondere:
Erweiterte Regelkreis-strukturen (Störgrößen-aufschaltung, Kaskaden-regelung, 2-DOF Design, Anti-Windup
Adaptive Control deals with the control of systems with unknown parameters. This is done by learning the controller parameters online and ensuring a stable operation of the closed-loop system. An adaptive controller, essentially, is a nonlinear controller that turns into a linear controller (for linear systems) once learning has finished. This lecture aims at introducing the theory and practice, abilities and limitations of the mainstream adaptive control concept, namely "Direct Model-Reference Adaptive Control" in continuous-time. This method is widely used e.g. in flight control of modern aircrafts. We will talk about other approaches and adaptive control of nonlinear systems, but focus on the understanding of the above, in theory and practice. The learning goals are how and why an adaptive controller works, when to use it, when not, how to improve it, and some robustness extensions.
The Machine Learning course first covers basic regression and classification methods (e.g. Bayesian Kernel Ridge Logistic Regression...) and then focuses on Bayesian formulations of learning (Bayes nets, probabilistic inference). In Stuttgart I plan to iterate the course every summer. Lectures are weekly, Thursdays, 14:00. Tutorials are weekly on Monday.
Learn how to achieve the following with psychtoolbox Drawing primitives, images, and textures Use alpha-mapping for blending and masking Play and loop sounds Accurate concurrent button-press detection Structuring experiments for reproducibility and robustness By reflecting upon the practical considerations that go into the implementation of psychological experiments, students will acquire a grasp of the different paradigms, technicalities and nuances of writing an experiment using Psychtoolbox.
Kinematics, Inverse Kinematics, Singularity, operational space, motion profiles, trajectory interpolation, Multiple tasks Dynamics, Reference Oscillator, PID, Euler-Lagrange equation, Newton-Euler recursion, Inverse and forward dynamics, Trajectory tracking, Operational space control Path Planning, trajectory optimization, Probabilistic Road Maps, RRTs, Non-holonomic systems Mobile Robotics, Simultaneous Localization and Mapping (SLAM) Control Theory, Hamilton-Jacobi-Bellman equation, LQ, Riccati, Controllability, Lyapunov and exponential stability, Reinforcement Learning in Robotics
Aufbauend auf den Inhalten von Einführung in die Regelungstechnik werden in dieser Vorlesung Konzepte aus folgenden Bereichen behandelt • Regelung nichtlinearer Systeme • Optimale Regelung • Robuste Regelung
In many applications and domains, massive amounts of data are collected and processed every day. To be able to make efficient use of such data, there is an urgent need for tools to extract important pieces of information from the flood of unimportant details. Machine learning is a relatively young discipline that tries to deal with this problem, by designing algorithms to analyze large amounts of complex data in a principled way. Machine learning is the core technique in many applications such as spam filtering, object recognition, analyzing user preferences, recommender systems, and so on. Scientific disciplines such as biology, neuroscience, physics, or medicine discover the potential of machine learning methods for analyzing their empirical data. And, last but not least, many large companies like google, Amazon, facebook heavily rely on machine learning techniques. The field of machine learning combines ingredients from several fields: we need to design efficient algorithms to process the amount of data, and we need to ensure that predictions made by machine learning algorithms are statistically sound. The focus of the lecture is on algorithmic and theoretical aspects of machine learning. We will cover many of the standard algorithms, learn about the general principles for building good machine learning algorithms, and analyze their theoretical properties. - Supervised learning problems: Linear methods; regularization; SVMS; kernel methods - Unsupervised learning problems: Dimension reduction (kernel PCA, multi-dimensional scaling, manifold methods); spectral clustering and spectral graph theory - How to model machine learning problems: Bayesian decision theory, loss functions, feature selection, evaluation and comparison of algorithms. Common pitfalls - Online algorithms - Learning theory (no free lunch theorem; generalization bounds; VC dimension; universal consistency; Theorem of Stone) - Low rank matrix methods (collaborative filtering, low rank matrix completion, compressed sensing) The following topics are NOT going to be covered: decision trees, neural networks / deep networks, graphical models, Bayesian approaches to machine learning, reinforcement learning.
Graphics processors contain hundreds of parallel processing elements and thus enable us to explore this realm of massively parallel computing today. The high number of parallel cores poses a great challenge for software design that must expose massive parallelism to benefit from the new hardware. The main purpose of the lecture is to teach practical algorithm design for such parallel hardware. Introduction to Parallel Computing Basic Algorithms: - Map, reduce, parallel branching, sorting - Parallel data storage and retrieval Parallel Computation: - FFT, particle systems - Parallel linear equation solvers, parallel PDEs - Parallel complexity analysis and profiling - System integration and graphics processor clusters
Linear Algebra, Vectors, dual vectors, coordinates, matrices, tensors, Coordinate free, The Singular Value Decomposition Theorem, Eigendecomposition, Power Method Derivatives, Coordinate free, total/partial, chain rules & autodiff ”Gradient vectors”, Co- and contra-variance, Taylor expansion Optimization, KKT conditions, Log Barrier, Augmented Lagrangian, The Lagrangian, unconstrained optimization, Backtracking line search, Wolfe conditions, & convergence, The Newton direction, Gauss-Newton, Quasi-Newton & BFGS, Conjugate Gradient, Blackbox & Global Optimization, Global Optimization as infinite bandidts Probabilities & Information, Bayes, Conjugate distributions, neg-log-probabilities & exp-neg-energies, Information, Entropie & Kullback-Leibler, The Laplace approximation, Variational Inference, The Fisher information metric, Maximum Entropy and Maximum Likelihood, Learning = Compression
- This lecture comprises different areas of Medical Data Science. Data Science or statistical machine learning methods have the potential to transform personal health care over the coming years. Advances in the technologies have generated large biological data sets. In order to gain insights that can then be used to improve preventive care or treatment of patients, these big data have to be stored in a way that enables fast querying of relevant characteristics of the data and consequently building statistical models that represent the dependencies between variables. These models can then be utilized to derive new biomedical principals, provide evidence for or against certain hypotheses, and to assist medical professionals in their decision process. Specific topics are: - Gaining new insights from medical data - Modeling uncertainty in medical data science models -Making medical findings available through interpretable decision support systems - Method-wise, the lecture will introduce methods for GWAS analyses (e.g., LMMs), methods for sequence analysis (e.g., kernel methods), methods for “small n problems” (e.g., domain adaptation, transfer learning, and multitask learning), methods for data integration (advanced unsupervised learning methods), methods for learning probabilistic Machine Learning models (e.g., graphical models), methods for large data sets (e.g., deep learning models)
Analyse und Synthese mehrschleifiger linearer Regelkreise im Zeit- und Frequenzbereich.
This lecture deals with Model Predictive Control (MPC), a modern control concept which has been actively researched and widely applied in industry in the last years. After an introduction to the basic ideas and stability concepts of MPC, more recent and current advances in research, like tube-based MPC considering robustness issues, economic MPC, distributed MPC, and stochastic MPC are discussed.
In recent years our experimental methods to record brain activity have been revolutionized. As the complexity of the data acquired by neurophysiologists increases, neural data analysis becomes ever more important: The complex multidimensional signals recorded with multielectrode arrays or two-photon imaging can no longer be interpreted by eye, but mathematical and statistical techniques are needed. In this practical course we will cover a selection of topics related to the analysis of different kinds of neural data: basic descriptive and inferential statistics, time series analysis, spike triggered average/covariance, spike sorting, dimensionality reduction techniques and information theory. The focus will be on hands-on experience in data analysis.
The course covers modern analysis and controller design methods for nonlinear systems:
In many practical control problems it is desired to optimize a given cost functional while satisfying constraints involving dynamical systems. These kind of problems typically fall into the area of optimal control, a centerpiece of modern control theory. This course gives an introduction to the theory and application of optimal control for linear and nonlinear systems. Topics covered in the course are: Nonlinear programming approach Dynamic programming Model predictive control Pontryagin maximum principle Applications
The Vision Sciences are an interdisciplinary field, with researchers having diverse backgrounds from psychology, biology, or medicine to physics, computer science and engineering. Analysing and designing experiments in the Vision Sciences thus requires knowledge straddling the typical boundaries of many disciplines. In this course we will cover some physics (light), electrical engineering (display devices), mathematical psychology (signal detection theory) and statistics (psychometric function estimation) in sufficient detail, to allow the students to both analyse and critically assess psychophysical experiments in the literature, as well as to design their own psychophysical experiments. Participants will acquire the necessary knowledge to critically assess experiments in the vision sciences as well as the necessary skills to design and analyse their own behavioural (psychophysical) experiments. Through homework assignments and computer exercises they will gain hands-on experience applying signal detection theory and psychometric function estimation to data, and avoid common pitfalls.
This is an advanced course on photo realistic image synthesis. The course will cover the theory of global illumination computations and will give an introduction to Monte Carlo and Quasi-Monte Carlo methods as one way of solving the rendering equation. In addition, we will discuss surface appearance models, and their representation based on spherical harmonics or wavelets, leading to other photo realistic rendering techniques such as precomputed radiance transfer methods.The goal is to achieve real-time global illumination.
Short description of the course: This seminar offers a unique possibility to glance at research questions, methods, and results in three research labs in Baden-Württemberg. The guiding question for the seminar is: How are biological and technical movements generated and controlled? In the seminar, you will get insights into this interdisciplinary field. We will give an overview including aspects like recording and analysis of human movement, reduced biomechanical models, neuro-musculo-skeletal models, humanoid robotics, control of complex human-like movements, and machine learning.
Lecture Psychophysical Methods
Seminar Spatial Vision
Seminar Colour Vision & Material Perception
Content - Loss functions and surrogate loss functions - Regularized empirical risk minimization - Ingredients from empirical process theory - Reproducing kernel Hilbert spaces - Kernel-based learning machines Level: M.Sc. course in mathematics Beginning this week, lectures are on: Monday 11:45 to 13:15, Friday 9:45 to 11:15 Beginning next week, the tutorials are on: Friday 14:00 to 15:30
Diese Lehrveranstaltung gibt eine Einführung in folgende Themengebiete. • Stochastische Simulation und Samplingverfahren • Bayessche Schätzverfahren und Filter • Gaußsche Prozesse und Regression
This course provides a rigorous introduction to stochastic calculus and stochastic processes. Anyone who is interested can participate.
The students are able to
Aufbauend auf Angewandte Statistik I werden komplexere statistische Methoden behandelt: Generalisierte Lineare Modelle (GLM), Hauptkomponentenanalyse (PCA), Unabhängigkeitsanalyse (ICA) und Bayes-Statistik. Der Schwerpunkt liegt auf der praktischen Anwendung aller Methoden und deren Implementation in der Programmiersprache Python (mit den Modulen statsmodels, scipy.stats, sklearn und pystan) und der Darstellung der Ergebnisse in Notebooks. Die Studenten sollen weiterführende statistische Methoden kennen-, anwenden und in Software implementieren lernen. Die Unterschiede zwischen frequentistischer und Bayes-Statistik werden hinterfragt. Angeeignetes Wissen und Erfahrung soll die Studenten in die Lage versetzen, Versuche selbst planen und auswerten zu können und dabei typische Fehler zu vermeiden. In der Literatur dargestellte Ergebnisse werden kritisch hinterfragt.
Wann immer in der Informatik Informationen visualisiert werden, kommen Bildschirme & Beamer zum Einsatz: In der Computergraphik, der Medizininformatik oder der Mensch-Computer-Interaktion. Bildschirme & Beamer werden auch bei visueller Stimulation in Experimenten der Kognitions- und den Neurowissenschaften verwendet, und für präzise Stimuluspräsentation müssen diese ausgemessen und kalibriert werden. Themen: Einführung in die Physik des Lichts und die Photometrie; Technik von Versuchsmonitoren (LCD,CRT, LED); Technik von optischen Messgeräten (Photometer, Spektrometer, Lichtmesskamera, Transientenrekorder); Vermessung und Optimieren der Darstellung; Programmieren präziser Experimente mit speziellen Toolboxen (Psychopy, Psychtoolbox). Studenten lernen die technischen Grundlagen von Bildschirmen & Beamern kennen. Das erworbene Verständnis versetzt die Studenten in die Lage, Vor- und Nachteile sowie Grenzen gängiger Display Technologien beurteilen und messen zu können. Am Ende der Vorlesung mit praktischen Übungen sind die Studierenden in der Lage, selbständig eigene visuelle Experimente in der Kognitionswissenschaft und der Medienwissenschaft durchzuführen und typische Fehler bei der Kalibrierung und Nutzung von Displays zu vermeiden.