All members of our the IMPRS-IS community are invited to attend our seventh annual interview symposium taking place from Tuesday, January 17, to Friday, January 20, 2023. The event will feature scientific talks from our applicants, three talks from our current Ph.D. Scholars, and two faculty keynotes by Dr. Daniel Häufle from the University of Tübingen and Dr. Mathias Niepert from the University of Stuttgart.
Each year in January, IMPRS-IS hosts an interview symposium to interview and recruit new Ph.D. students. We are thrilled to announce that 144 applicants have accepted our invitation to join for the interview symposium in hopes of securing a Ph.D. position within our program.
The four-day event will feature scientific talks from the applicants, as well as two keynotes by our faculty members Dr. Daniel Häufle and Dr. Mathias Niepert. This year we have also added three scientific talks from our current scholars as part of the event's welcome.
This event is closed to the general public, but all members of the IMPRS-IS community are welcome to join for the talks. For details about how to participate, please email us (firstname.lastname@example.org). Here are the details about the event keynotes and scholar talks.
Date: Tuesday, January 17, 2023
Time: 15:15 - 15:30
Title: Haptify: a measurement-based benchmarking system for grounded force-feedback devices
Abstract: Grounded force-feedback (GFF) devices are an established and diverse class of haptic technology based on robotic arms. However, the number of designs and how they are specified make comparing devices difficult. We thus present Haptify, a benchmarking system that can thoroughly, fairly, and noninvasively evaluate GFF haptic devices. The user holds the instrumented device end-effector and moves it through a series of passive and active experiments. Haptify records the interaction between the hand, device, and ground with a seven-camera optical motion-capture system, a 60-cm-square custom force plate, and a customized sensing end-effector. We demonstrate six key ways to assess GFF device performance: workspace shape, global free-space forces, global free-space vibrations, local dynamic forces and torques, frictionless surface rendering, and stiffness rendering. We then use Haptify to benchmark two commercial haptic devices. With a smaller workspace than the 3D Systems Touch, the more expensive Touch X outputs smaller free-space forces and vibrations, smaller and more predictable dynamic forces and torques, and higher-quality renderings of a frictionless surface and high stiffness.
Biography: Farimah Fazlollahi received the B.S. degree in mechanical engineering with a major in applied design and the M.S degree in mechanical engineering with a major in applied design, dynamics, vibrations, and control from Shiraz University, Shiraz, Iran, in 2015 and 2017, respectively. She is currently working toward the Ph.D. degree in mechanical engineering with the Haptic Intelligence Department, Max Planck Institute for Intelligent Systems, Stuttgart, Germany. Her research interests include haptics, robotics, dynamics, mechanism design, and sensor fusion.
Date: Tuesday, January 17, 2023
Time: 15:30 - 15:45
Title: Facilitating generalization for representation learning under distribution shifts
Abstract: Successful deployment of machine learning models requires generalization beyond the training distribution, where data at test time can experience varying types of (out-of-) distribution shifts. Consequently, this presentation will cover two works from our lab that investigate how one can improve generalization in representation learning under semantic out-of-distribution shifts using normalizing flows, and how one can successfully adapt well-generalizing foundation models under continuous distribution shifts.
In this talk, I will discuss two ways to formalize dynamic decision making problems. One, called performative prediction, directly makes assumptions about the aggregate population response to a decision rule. The other, called strategic classification, follows microeconomic tradition in modeling individuals as utility-maximizing agents with perfect information. I will reflect on the advantages and limitations of either perspective, pointing out avenues for future research.
Based on collaborations with Anca Dragan, Meena Jagadeesan, Celestine Mendler-Dünner, John Miller, Smitha Milli, Juan Carlos Perdomo, Tijana Zrnic
Biography: Karsten Roth is a Ph.D. researcher at the Explainable Machine Learning group as part of the IMPRS-IS and the ELLIS program, co-supervised by Zeynep Akata at the University of Tuebingen and Oriol Vinyals at Deepmind. Karsten has completed both Bachelor and Master studies in Physics at Heidelberg University (2021). He has spent time abroad in Canada as a research intern at the Montreal Institute for Learning Algorithms (MILA) with Joseph Paul Cohen and Yoshua Bengio, and the Vector Institute with Marzyeh Ghassemi, working on all manners of representation learning and their applications to the medical domain. As research intern, Karsten has also worked at the Amazon AWS research lablet in Tuebingen on Anomaly Detection with Peter Gehler and Thomas Brox, and Meta AI in Paris on Disentangled Representation Learning with Pascal Vincent and Diane Bouchacourt. His primary interests cover the study of generalisation properties of modern representation learning methods, with applications to zeroshot, fewshot and continual learning problems, as well as their application to medicine and the sciences.
Date: Tuesday, January 17, 2023
Time: 15:45 - 16:00
Title: Curious exploration via structured world models
Abstract: It has been a long-standing dream to design artificial agents that explore their environment efficiently via intrinsic motivation, similar to how children perform curious free play. Despite recent advances in intrinsically motivated reinforcement learning (RL), sample-efficient exploration in object manipulation scenarios remains a significant challenge as most of the relevant information lies in the sparse agent-object and object-object interactions. In this talk, I will discuss our recent work using structured world models to incorporate relational inductive biases in the control loop to achieve sample-efficient and interaction-rich exploration in compositional multi-object environments. I will also showcase how the self-reinforcing cycle between good models and good exploration opens up another avenue: zero-shot generalization to downstream tasks via model-based planning.
Biography: Cansu Sancaktar is currently a PhD student at the MPI for Intelligent Systems in the Autonomous Learning group, led by Georg Martius, and joined IMPRS-IS in April 2021. She completed her Bachelor’s and Master’s degrees in Electrical Engineering and Information Technology at the Technical University of Munich as a DAAD and Max Weber Program scholar. In her PhD, she is focusing on unsupervised reinforcement learning, where the goal is to achieve sample-efficient exploration via child-like free play in artificial agents.
Dr. Daniel Häufle
Date: Wednesday, January 18, 2023
Time: 16:45 - 17:30
Title: Interacting with AI
Abstract: Currently to me, the biggest thrill in robotics is to see that robots learn to interact with the real world. Robots are now able to walk in uncertain environments and deal with perturbations. While this is exciting to study and to see, I would claim that biology solved these problems millions of years ago. Surprisingly, we still do not fully understand the principles behind the biological solution. I have the great pleasure to collaborate with fantastic colleagues in the IMPRS-IS. Together, we try to better understand the interaction between neuronal circuits, musculoskeletal dynamics, and the environment. In my talk, I will present past and ongoing research within IMPRS-IS. I will show why and how we develop computer simulations of neuro-muscular control and learning, why and how we translate them into robotic concepts, and why I believe that this is relevant for robotic assistance in neuro-rehabilitation.
Biography: Daniel Häufle is assistant professor and head of the Research Group Multi-Level Modeling in Motor Control and Rehabilitation Robotics at Hertie Institute for Clinical Brain Research, University of Tübingen, and the Institute for Modelling and Simulation of Biomechanical Systems, University of Stuttgart. His group investigates how muscles and the nervous system interact to generate human movement and how this interaction may be impaired in neurological movement disorders. He studied physics and biomechanics in Jena, Germany and Calgary, Canada. For his PhD, he worked in the Computational Biopyhsics & Biorobotics Lab of Syn Schmitt at the University of Stuttgart. With a Fulbright Scholarship he visited the Robotics Institute of Carnegie Mellon University in Pittsburgh, USA. His habilitation in Computer Science at the University of Tübingen focused on the contribution of morphology to the control of biological movement.
Dr. Mathias Niepert
Date: Thursday, January, 20, 2022
16:45 - 17:15
Data, decisions, and dynamics
Abstract: Machine learning at scale has led to impressive results ranging from text-based image generation, reasoning with natural language, and code synthesis to name but a few. ML at scale is also successfully applied to a broad range of problems in engineering and the sciences. These recent developments make some of us question the utility of incorporating prior knowledge in form of symbolic (discrete) structures and algorithms. Is computing and data at scale all we need?
We will make an argument that discrete (symbolic) structures and algorithms in machine learning models are advantageous and even required in numerous application domains such as Biology, the Material Sciences, and Physics. Biomedical entities and their structural properties, for example, can be represented as graphs and require inductive biases equivariant to certain group operations. My labs research is concerned with the development of machine learning methods that combine discrete structures with continuous equivariant representations. We also address the problem of learning and leveraging structure from data where it is missing, combining discrete algorithms and probabilistic models with gradient-based learning. We will show that discrete structures and algorithms appear in numerous places such as ML-based PDE solvers and that modeling them explicitly is indeed beneficial. Especially machine learning models with the aim to exhibit some form of explanatory properties have to rely on symbolic representations. The talk will also cover some biomedical and physics related applications.
Biography: Mathias Niepert is a professor at the University of Stuttgart and a faculty member of the International Max Planck Research School for Intelligent Systems (IMPRS-IS). He heads the Machine Learning and Simulation Lab. His professorship is part of the Cluster of Excellence for the Simulation Sciences (SimTech), the Department of Computer Science, and the ELLIS society. He is also a Chief Scientific Advisor at NEC Laboratories Europe. At NEC Labs Europe he was senior (2015-2017) and chief research scientist (2017-2021) as well as manager (2019-2021) of the machine learning group. From 2013-2015 he was a postdoctoral research associate at the Allen School of Computer Science, University of Washington, Seattle. Dr. Niepert obtained his PhD from Indiana University. His group's research interests include representation learning for discrete structures and algorithms, geometric deep learning, probabilistic graphical models, and the intersection of ML and the sciences.
For access to these talks, please contact Sara Sorce (email@example.com).
Photo credit: MPI für Intelligent Systeme