All current and former employees and friends of the Max Planck Institute for Intelligent Systems and Cyber Valley are welcome to attend this event. If you have any questions, please contact Cyber Valley Research Coordinator - Florian Mayer (email@example.com)
|Wednesday January 12|
|12:45-1:00 pm||Opening remarks|
|1:00-2:30 pm||Presentations | Part One|
|3:00-4:30 pm||Presentations | Part Two|
|Thursday January 13|
|8:45-9:00 am||Opening remarks|
|9:00-10:30 am||Presentations | Part Three|
|11:00 am - 12:30 pm||Presentations | Part Four|
Towards Machines that See the World like Humans
Wieland Brendel (University of Tübingen, Germany)
Talk: Wednesday, January 12, 2021 – 1:00-1:30 pm
In the past few years, deep neural networks have surpassed human performance on a range of complex cognitive tasks. However, unlike humans, these models can be derailed by almost imperceptible perturbations, often fail to generalize beyond the training data and require large amounts of data to learn novel tasks. A core reason for this behavior is shortcut learning, i.e. the tendency of neural networks to pick up statistical signatures sufficient to solve a given task instead of learning the underlying causal structures and mechanisms in the data. My research ties together adversarial machine learning, disentanglement, interpretability, self-supervised learning, and theoretical approaches like nonlinear Independent Component Analysis to develop theoretically grounded yet empirically successful visual representation learning techniques that can uncover the underlying structure of our visual world and close the gap between human and machine vision.
An Algorithmic Safety View of Learning in Human-centered Domains
Shalmali Joshi (Harvard University, Cambridge, USA)
Talk: Wednesday, January 12, 2021 – 1:30-2:00 pm
Machine Learning advances have revolutionized many domains such as machine translation, complex game playing, and scientific discovery. On the other hand, ML has only enjoyed modest successes in human-centered applications. To improve the utility, reliability, and robustness of Machine Learning (ML) models in human-centered domains, we need to address several foundational challenges. In this talk, I will demonstrate how an algorithmic-safety perspective can motivate specific technical challenges for learning in human-centered domains such as healthcare. Specifically, I will discuss the need to improve the utility of ML-robustness, explainability with an emphasis on decision-making, and post-hoc algorithmic safety to prevent harm. I will discuss my contributions on i) novel methods to improve causal robustness of ML methods designed for practical generative settings, ii) aiding safe decision-making in non-IID settings using time-series explainability intended to address clinicians’ requirements, and iii) novel learning algorithms to optimize for post-deployment safety in sequential decision-making settings. I will conclude with an overview of my future research vision on novel safety-based objectives for explainability in ML, expanding ML-based solutions to general and practical generative settings, and outlining novel ways of validating ML models targeting safety-based objectives.
On Inferential Challenges in Modern Machine Learning
Jiaxin Shi (Microsoft Research New England, Cambridge, USA)
Talk: Wednesday, January 13, 2021 – 2:00-2:30 pm
Machine learning (ML) can be viewed as solving two types of inferential tasks: inferring models from data and inferring knowledge from models. Central to this perspective is the development of fast, accurate inference methods for model parameters and latent variables. In this talk, I will present my work on addressing inferential challenges in modern ML which are characterized by enhanced model complexity and massive datasets that demand algorithms at scale. I will focus on 1) theory and methods of score estimation, an emerging tool for dealing with intractable probability densities in today's highly expressive ML models, and 2) orthogonal inducing points, an inferential idea that advanced the state-of-the-art of scalable uncertainty models in ML. Finally, I will discuss future directions that connect these ideas to generative modeling, continual learning, and how they could help us understand deep learning.
Computer Vision for Global-scale Biodiversity Monitoring
Sara Beery (California Institute of Technology, Pasadena, USA)
Talk: Wednesday, January 12, 2021 – 3:00-3:30 pm
We require a real-time, modular earth observation system that unites efforts across research groups in order to provide the vital information necessary for global-scale impact in sustainability and conservation in the face of climate change. The development of such systems requires collaborative, interdisciplinary approaches that translate diverse sources of raw information into accessible scientific insight. For example, we need to monitor species in real time and in greater detail to quickly understand which conservation efforts are most effective and take corrective action. Current ecological monitoring systems generate data far faster than researchers can analyze it, making scaling up impossible without automated data processing. However, ecological data collected in the field presents a number of challenges that current methods, like deep learning, are not designed to tackle. These include strong spatiotemporal correlations, imperfect data quality, fine-grained categories, and long-tailed distributions. My work seeks to overcome these challenges, and includes methods which can learn from imperfect data, systematic frameworks for measuring and overcoming performance drops due to domain shift, and the deployment of efficient human-AI systems that have made significant real-world conservation impact. My future research agenda will expand upon the strong foundation built by my past and current research. It will seek to make effective use of all available modalities of data, incorporate expert knowledge systematically, and ensure these systems are equitable and ethical – all fundamental and unresolved challenges for CV&ML.
Open-World Learning for Biomedicine
Maria Brbic (Stanford University, Stanford, USA)
Talk: Wednesday, January 13, 2021 – 3:30-4:00 pm
Biomedical data poses multiple hard challenges that break conventional machine learning assumptions. In this talk, I will highlight the need to move beyond our prevalent machine learning paradigms and methods to enable them to drive novel biomedical discoveries. I will focus on open-world deep learning methods that generalize to new scenarios never seen during training and demonstrate the impact they have in single-cell genomics. I will first present a method that transfers knowledge across heterogeneous datasets generated under different distributions, and then describe the paradigm needed to discover previously unobserved phenomena. I will discuss the biological findings enabled by these methods and the conceptual shift they bring in annotating comprehensive single-cell atlas datasets. Altogether, my work demonstrates that generalization to never-before-seen scenarios is not only possible, but it is a necessary component in developing next-generation methods that can reveal new scientific insights.
Social Robots for Health and Education
Patrícia Alves-Oliveira (University of Washington, Seattle, USA)
Talk: Wednesday, January 12, 2021 – 4:00-4:30 pm
I build innovative social robots that address important challenges in health care and education. A unifying challenge that I address in my work is that most of the existing resources that treat complex health and learning problems are developed for professionals. My vision is to democratize health care and education by using social robots to break barriers of access to these critical domains. Towards this, I designed, developed, fabricated, and deployed robots capable of reciprocal interactions with humans to be used as valuable sources of knowledge and assistance that empower people to make decisions about their mental health, social-emotional learning, and creativity. My work probes the intersection of the fields of Human-Robot Interaction, Human-Centered Design, and Artificial Intelligence as essential building blocks for creating social robots that can empower human health and education.
Intelligent Soft Systems – How Fundamental Physical Principles Enable Function
Philipp Rothemund (Max Planck Institute for Intelligent Systems, Stuttgart, Germany)
Talk: Thuradsy, January 13, 2020 – 9:00-9:30 am
Functional soft materials promise to revolutionize robotics hardware by providing entirely new functions and by allowing robots to perform better than traditional robots made from materials such as metals and electromagnetic motors. Over the last decade a large variety of intelligent systems based on functional soft materials have been successfully demonstrated. However, only the surface of the capabilities of these new types of materials has been skimmed, because the fundamental materials behavior, from which they derive their functionality, is often not sufficiently understood.
In this talk, I discuss fundamental physical principles of soft materials, which can enable function in high-performing soft devices. First, I describe how snap-through instabilities in elastomeric structures can be exploited for electronics-free control of robots. Then, I discuss how the electromechanical coupling between electric fields and soft structures enables electroactive devices and artificial muscles. Finally, I outline nonlinear and thermomechanical effects, which have the potential to serve as the basis for the next generation of intelligent soft systems.
Advanced Multimodal Locomotion and Multifunctionality of Soft-bodied Millirobots
Wenqi Hu (Max Planck Institute for Intelligent Systems, Stuttgart, Germany)
Talk: Thursday, January 13, 2020 – 9:30-10:00 am
Large individually-controlled degrees of freedom (DOF) in animal epidermises enable the animals to dynamically and dexterously interact with their natural environment for diverse functions. This capability still holds at small length scales: flatworms, starfish larvae, corals, comb jellyfishes, and millipedes utilize their numerous individually-controlled cilia, tube feet, or tiny legs for highly adaptable locomotion and multiple functions, such as nutrient transportation, predation, and bio-mixing. Despite their significant future potential applications and numerous recent advances, current miniature (centi-/millimeter) robots still do not have similar tiny cilia, feet, or legs in their epidermises to dynamically and dexterously interact with their operation environment, which severely limits their locomotion capabilities and functionalities.
The presentation will first walk through my previous projects of using physical intelligence, such as soft design, flow structure interaction, and metachronal coordination, to enhance miniature robots' locomotion capability and functions. Then, I will briefly discuss how I support these investigations on robotics by innovating on micro-fabrication and novel materials. At last, I will propose a new methodology to address the prementioned scientific challenge directly. Such an approach would immediately augment the functional interaction of various miniature robots with their surrounding environments, leading to disruptive locomotion and manipulation capabilities for various future applications.
Structured Learning – Encoding Structure from Physics, Algorithms, and Geometry into Machine Learning for Robotics and Computer Vision
Rico Jonschkowski (Google Brain, Mountain View, USA)
Talk: Thursday, January 13, 2020 – 10:00-10:30 am
This talk focuses on the interaction of structural assumptions and machine learning. While modern machine learning (aka deep learning) is able to leverage large amounts of data and computation to achieve important results, I believe that the recent shift of the community towards ever more general learning approaches is flawed. The progress of deep learning so far has relied heavily on advances in computation, data collection, and software libraries. But as such improvements are plateauing, future progress in AI requires us to reevaluate the role of structure in learning and to tightly integrate both for better data efficiency, generalization, and explainability. In this talk, I will present how structure from physics, algorithms, and geometry can be encoded in learning approaches to robotics and computer vision and why we need a better fundamental understanding of the interaction of structure, learning, and optimization.
Towards Democratic Learning and Collective Intelligence
Krikamol Muandet (Max Planck Institute for Intelligent Systems, Tübingen, Germany)
Talk: Thursday, January 13, 2020 – 11:00-11:30 am
An ability to generalize to unseen environments is of paramount importance in machine learning. In order to equip artificially intelligent systems with such capability, it is necessary to deal with various forms of distributional shifts and to distinguish between spurious correlations and genuine ones in the observed data. To this end, the first part of this talk will be about the Hilbert space embedding of distributions, or kernel mean embedding (KME) for short, which is a nonparametric kernel-based framework to represent probability distributions and model changes thereof. In particular, I will focus on how this framework can help improve the credibility of algorithmic decision making based on observational data by enabling us to reason about higher-order causal effects of policy interventions as well as by removing the effect of unobserved confounders through the use of an instrumental variable (IV) and a proxy variable. In the second part of this talk, we will take a sober look at the foundation of learning and generalization. In order to build intelligent systems that can truly generalize to the real world, we must take a step back and question the traditional definition of generalization and standard procedures in machine learning. Since society is made up of a set of diverse individuals, demographic groups, and institutions, learning and deploying algorithmic models across heterogeneous environments face a set of various trade-offs. I will argue that in such heterogeneous environments, creating machine learning frameworks that allow for real-world generalization is reminiscent of designing algorithms that respect the democratic principle.
Diversity and Inequality in Social Networks
Ana-Andreea Stoica (Columbia University, New York, USA)
Talk: Thursday, January 13, 2020 – 11:30 am - 12:00 pm
Online social networks often mirror inequality in real-world networks, from historical prejudice, economic or social factors. Such disparities are often picked up and amplified by algorithms that leverage social data for the purpose of providing recommendations, diffusing information, or forming groups. In this talk, I discuss an overview of my research involving explanations for algorithmic bias in social networks, briefly describing my work in information diffusion, grouping, and general definitions of inequality. Using network models that reproduce inequality seen in online networks, we'll characterize the relationship between pre-existing bias and algorithms in creating inequality, discussing different algorithmic solutions for mitigating bias.
Towards Principled Algorithmic Support of Human Decision-Making
Ivan Stelmakh (Carnegie Mellon University, Pittsburgh, USA)
Talk: Thursday, January 13, 2020 – 12:00-12:30 pm
Many important applications such as hiring, healthcare, and scientific peer review rely on human decision-making. Recently, the scale of many of these applications has increased dramatically which is both an opportunity and a challenge. On the one hand, the large amount of data generated in these applications opens up an opportunity to take a novel data-centric perspective on the classical problems of human decision-making. On the other hand, the large scale makes it hard or even impossible for humans to do all the work manually; hence, there is a challenge of developing principled algorithmic tools to support human decision-makers. In this talk, I will discuss my work on exploring the opportunities and addressing the challenges in the context of scientific peer review. In that, I will talk about empirical and theoretical work that has impacted major Computer Science conferences such as NeurIPS and ICML. I will then outline ideas for future work.