Institute Homepage
Institute Homepage DE Sign In

Publications

DEPARTMENTS

Emperical Interference

Haptic Intelligence

Modern Magnetic Systems

Perceiving Systems

Physical Intelligence

Robotic Materials

Social Foundations of Computation

Research Groups

Autonomous Vision

Autonomous Learning

Bioinspired Autonomous Miniature Robots

Dynamic Locomotion

Embodied Vision

Human Aspects of Machine Learning

Intelligent Control Systems

Learning and Dynamical Systems

Locomotion in Biorobotic and Somatic Systems

Micro, Nano, and Molecular Systems

Movement Generation and Control

Neural Capture and Synthesis

Physics for Inference and Optimization

Organizational Leadership and Diversity

Probabilistic Learning Group

Topics

Robot Learning

Conference Paper

2022

Autonomous Learning

Robotics

AI

Career

Award

Article Uncertainty in learning, choice and visual fixation Stojić, H., Orquin, J., Dayan, P., Dolan, R., Speekenbrink, M. {Proceedings of the National Academy of Sciences of the United States of America}, 117(6):3291-3300, National Academy of Sciences, Washington, D.C., 2020
{Uncertainty plays a critical role in reinforcement learning and decision making. However, exactly how it influences behavior remains unclear. Multiarmed-bandit tasks offer an ideal test bed, since computational tools such as approximate Kalman filters can closely characterize the interplay between trial-by-trial values, uncertainty, learning, and choice. To gain additional insight into learning and choice processes, we obtained data from subjects\textquoteright overt allocation of gaze. The estimated value and estimation uncertainty of options influenced what subjects looked at before choosing; these same quantities also influenced choice, as additionally did fixation itself. A momentary measure of uncertainty in the form of absolute prediction errors determined how long participants looked at the obtained outcomes. These findings affirm the importance of uncertainty in multiple facets of behavior and help delineate its effects on decision making.}
DOI BibTeX
Miscellaneous Uncovering the organization of neural circuits with generalized phase locking analysis Safavi, S., Panagiotaropoulos, T., Kapoor, V., Ramirez-Villegas, J., Logothetis, N., Besserve, M. 2020
{Spike-field coupling characterizes the relationship between neurophysiological activities observed at two different scales: on the one hand, the action potential produced by a neuron, on the other hand a mesoscopic \textquotedblleftfield\textquotedblright signal, reflecting subthreshold activities. This provides insights about the role of a specific unit in network dynamics. However, assessing the overall organization of neural circuits based on multivariate data requires going beyond pairwise approaches, and remains largely unaddressed. We develop Generalized Phase Locking Analysis (GPLA) as an multichannel extension of univariate spike-field coupling. GPLA estimates the dominant spatio-temporal distributions of field activity and neural ensembles, and the strength of the coupling between them. We demonstrate the statistical benefits and interpretability of this approach in various biophysical neuronal network models and Utah array recordings. In particular, we show that GPLA, combined with neural field modeling, help untangle the contribution of recurrent interactions to the spatio-temporal dynamics observed in multi-channel recordings.}
DOI BibTeX
Miscellaneous Uncovering the organization of neural circuits with generalized phaselocking analysis Safavi, S., Panagiotaropoulos, F., Kapoor, V., Ramirez-Villegas, J., Logothetis, N., Besserve, M. Computational and Systems Neuroscience Meeting (COSYNE 2020), 150-151, 2020
{Many neurophysiological recordings provide concurrent signals of two different nature: On the one hand, the timestamps of action potentials reflect the information sent by individual neurons, and on the other hand, Local field Potentials (LFP) reflect multiple sub-threshold and postsynaptic mechanisms related to the underlying network activity. The synchronization between spiking activity and the phase of particular LFP rhythms has been used as an important marker to reason about the underlying cooperative network mechanisms. In order to extract in a systematic way coupling information from the largely multivariate data available in current recording techniques, we study a multivariate extension of spike-field coupling analysis. After whitening band-passed LFP signals, we collect normalized pairwise complex-value spike-field coupling coefficients in a rectangular matrix and summarize its structure with the largest singular value and the corresponding singular vectors. Singular vectors represent the dominant LFP and spiking patterns and the singular value, called generalized Phase Locking Value (gPLV), characterizes the strength of the coupling between LFP and spike patterns. We further investigate the statistical properties of the gPLV and develop an empirical and theoretical statistical testing framework. We apply the method to various simulated and experimental datasets. first, GPLA\textquoterights performance is superior to univariate measures in the presence of large amounts of noise. Next, application of GPLA on simulations of hippocampal sharp-waveripples (SWR) reveals various characteristics of hippocampal circuitry (e.g. communication flow from CA3 to CA1 during the SWR) with minimal prior knowledge. Furthermore, application of GPLA on Utah array recordings in anesthetized macaque suggests a non-trivial coupling between spiking activity and LFP traveling wave in the ventrolateral Prefrontal Cortex (vlPFC). In summary, with GPLA, we can quantify, characterize and statistically assess the interactions between population spiking activity and mesoscopic network dynamics.}
BibTeX
Article Unexpected Trends in the Stability and Dissociation Kinetics of Lanthanide(III) Complexes with Cyclen-Based Ligands across the Lanthanide Series Garda, Z., Nagy, V., Rodr\’\iguez-Rodr\’\iguez, A., Pujales-Paradela, R., Patinec, V., Angelovski, G., Tóth, E., Kálmán, F., Esteban-Gómez, D., Tripier, R., Platas-Iglesias, C., Tircsó, G. {Inorganic Chemistry}, 59(12):8184-8195, American Chemical Society, Washington, DC, 2020
{We report a detailed study of the thermodynamic stability and dissociation kinetics of lanthanide complexes with two ligands containing a cyclen unit, a methyl group, a picolinate arm, and two acetate pendant arms linked to two nitrogen atoms of the macrocycle in either cis (1,4-H3DO2APA) or trans (1,7-H3DO2APA) positions. The stability constants of the Gd3+ complexes with these two ligands are very similar, with log KGdL values of 16.98 and 16.33 for the complexes of 1,4-H3DO2APA and 1,7-H3DO2APA, respectively. The stability constants of complexes with 1,4-H3DO2APA follow the usual trend, increasing from log KLaL \textequals 15.96 to log KLuL \textequals 19.21. However, the stability of [Ln(1,7-DO2APA)] complexes decreases from log K \textequals 16.33 for Gd3+ to 14.24 for Lu3+. The acid-catalyzed dissociation rates of the Gd3+ complexes differ by a factor of $\sim$15, with rate constants (k1) of 1.42 and 23.5 M-1 s-1 for [Gd(1,4-DO2APA)] and [Gd(1,7-DO2APA)], respectively. This difference is magnified across the lanthanide series to reach a 5 orders of magnitude higher k1 for [Yb(1,7-DO2APA)] (1475 M-1 s-1) than for [Yb(1,4-DO2APA)] (5.79 $\times$ 10-3 M-1 s-1). The acid-catalyzed mechanism involves the protonation of a carboxylate group, followed by a cascade of proton-transfer events that result in the protonation of a nitrogen atom of the cyclen unit. Density functional theory calculations suggest a correlation between the strength of the Ln-Ocarboxylate bonds and the kinetic inertness of the complex, with stronger bonds providing more inert complexes. The 1H NMR resonance of the coordinated water molecule in the [Yb(1,7-DO2APA)] complex at 176 ppm provides a sizable chemical exchange saturation transfer effect thanks to a slow water exchange rate of (15.9 $\pm$ 1.6) $\times$ 103 s-1.}
DOI BibTeX
Miscellaneous Universal Parallel Transmit Pulse Design for 3-D Local-Excitation based on different sized databases of B0/B1+-maps: A 7T Study Geldschläger, O., Shao, T., Herrler, J., Nagel, A., Henning, A. 2020 ISMRM & SMRT Virtual Conference & Exhibition, 2020
{This study investigates universal parallel-transmission (pTx) radio-frequency-pulses for 3-dimensional local-excitation designed on different sized databases of B0/B1+-maps from human heads at 7T. Thus, it prospectively abandons the need for time-consuming subject specific B0/B1+-mapping and pTx-pulse calculation during the scan session. For the proposed calculation routine, the design-database does not need to include more than five heads, to achieve a pTx-pulse that excites the same 3-dimensional local-excitation target-pattern on the tested 40 different heads. The resulting universal pulses created magnetization-profiles with (in most cases) an only marginally worse Normalized-Root-Mean-Square-Error compared to the magnetization-profiles produced by pulses tailored to individual heads.}
BibTeX
Miscellaneous V1 serves as a motor cortex for visually guided saccades: Evidence from human and monkey behavioral and neural data Zhaoping, L. 2020 BibTeX
Article Visual appearance modulates motor control in social interactions de la Rosa, S., Meilinger, T., Streuber, S., Saulton, A., Fademrecht, L., Quiros-Ramirez, M., Bülthoff, H., Bülthoff, I., Ca nal-Bruland, R. {Acta Psychologica}, 210:1-9, Elsevier, Amsterdam, 2020
{The goal of new adaptive technologies is to allow humans to interact with technical devices, such as robots, in natural ways akin to human interaction. Essential for achieving this goal, is the understanding of the factors that support natural interaction. Here, we examined whether human motor control is linked to the visual appearance of the interaction partner. Motor control theories consider kinematic-related information but not visual appearance as important for the control of motor movements (Flash \& Hogan, 1985; Harris \& Wolpert, 1998; Viviani \& Terzuolo, 1982). We investigated the sensitivity of motor control to visual appearance during the execution of a social interaction, i.e. a high-five. In a novel mixed reality setup participants executed a high-five with a three-dimensional life-size human- or a robot-looking avatar. Our results demonstrate that movement trajectories and adjustments to perturbations depended on the visual appearance of the avatar despite both avatars carrying out identical movements. Moreover, two well-known motor theories (minimum jerk, two-thirds power law) better predict robot than human interaction trajectories. The dependence of motor control on the human likeness of the interaction partner suggests that different motor control principles might be at work in object and human directed interactions.}
DOI BibTeX
Miscellaneous Visual illusions and feedforward and feedback processes in visual reccognition Zhaoping, L. Bernstein Conference 2020, 2020
{Feedback neural connections are abundant from higher to lower visual cortical areas. Although they are manifested in visual recognition behavior (e.g., Tang et al 2018) and even exploited in artificial convolutional neural networks (Cao et al , 2015) which dominantly use feedforward processes only, the recurrent nature of the processing makes it difficult to understand how they work together with the feedforward visual processing. Motivated by Zhaoping\textquoterights proposal to study the feedback in the context of the attentional bottleneck and her hypothesized central-peripheral dichotomy (Zhaoping 2019) that the top-down feedback to aid visual recognition, using the computation of analysis-by-synthesis, is stronger in central than peripheral visual field, and taking advantage of our better knowledge about the feedforward signals from the primary visual cortex (V1), we investigate this topic combining computational and visual psychophysical means. Computationally, we illustrate how the same activities of the V1 neurons tuned to, e.g., orientation, motion direction, or stereo depth, could arise from very different kinds of visual inputs (as observed physiologically, e.g., Cummings and Parker 1997, Kuriki et al 2008). These different kinds of inputs that evoke the same V1 responses could therefore cause visual perceptual confounds or visual illusions. Some visual illusions (e.g., the reversed phi motion) in the literature can indeed be understood accordingly, as have also been noted by previous researchers. Since top-down feedback can combine internal knowledge about the visual world and feedforward inputs to disambiguate between the visual confounds, we ask whether the central-peripheral dichotomy could explain when V1 activities do or do not give rise to visual percepts or illusions, how the strengths of these percepts depend on whether the visual inputs are viewed in the central and peripheral visual field and on whether the visual inputs are presented long enough or too briefly to make the feedback more or less effective. To this end, we explore what new illusions can be predicted by the central-peripheral dichotomy and by our knowledge about V1. We will present visual psychophysical tests of some of the predictions, and relate our findings to previous works such as those in feature finding, visual masking, and adversarial attacks in artificial neural networks.}
DOI BibTeX
Article What are we curious about? Brändle, F., Wu, C., Schulz, E. {Trends in Cognitive Sciences}, 24(9):685-687, Elsevier Current Trends, Kidlington, Oxford, UK, 2020
{What are we curious about? Dubey and Griffiths propose a rational theory of curiosity that unifies previously contradictory novelty-based and complexity accounts. It also paves the way for future investigations, such as studying approximate models of curiosity as well as what causes abnormal levels of exploration.}
DOI BibTeX
Miscellaneous What percentage of dots in this image are black? Visual saliency may distort perceived summary statistics Badler, J., Zhaoping, L. Twentieth Annual Meeting of the Vision Sciences Society (VSS 2020), 2020
{Human perception of probability is systematically distorted, with low probabilities typically overestimated and high probabilities underestimated. The distortion is thought to arise from constraints in the internal representations of probabilities (Zhang, Ren \& Maloney, 2019, bioRxiv). Here we present data suggesting that for visual stimuli, saliency may provide an additional contribution. In a visual scene, infrequent objects among frequent background objects tend to be more salient to attract attention, which in turn weights them more heavily in probability estimations. For example, if a gray visual field contains a few black dots among many more white dots of equal size and contrast, the black dots are more salient to attract attention. Thus, the proportion of dots that are black may appear larger than it actually is. Analogously, when the black dots are more numerous than the white dots, they become less salient than the white dots and their proportion can thereby appear smaller than it actually is. Since saliency effects are stronger peripherally (Zhaoping 2014, Oxford University Press), we predict that the probability distortion should be larger when objects are distributed in a larger field. This is because in a large field, more dots fall in the peripheral field for any given fixation location, making the probability estimation more reliant on the peripheral estimations of the summary statistics of the field. To test this, we implemented a variant of the task first reported by Zhang and Maloney (2012, Frontiers in Neuroscience). Four participants (three na\"\ive) viewed a square field containing a mixture of black and white dots (diameter\textequals0.26\mbox{$^\circ$}, density\textequals2/deg\mbox{$^2$}) subtending either 10\mbox{$^\circ$} or 20\mbox{$^\circ$}. After freely viewing for 1.5 seconds, participants had to estimate the fraction of dots that were black. For all participants, probability distortions were greater for the large field.}
BibTeX
Article Whole brain snapshot CEST at 3T using 3D-EPI: Aiming for speed, volume, and homogeneity Mueller, S., Stirnberg, R., Akbey, S., Ehses, P., Scheffler, K., Stöcker, T., Zaiss, M. {Magnetic Resonance in Medicine}, 84(5):2469-2483, Wiley-Liss, New York, 2020
{PURPOSE: CEST MRI enables imaging of distributions of low-concentrated metabolites as well as proteins and peptides and their alterations in diseases. CEST examinations often suffer from low spatial resolution, long acquisition times, and concomitant motion artifacts. This work aims to maximize both resolution and volume coverage with a 3D-EPI snapshot CEST approach at 3T, allowing for fast and robust whole-brain CEST MRI. METHODS: Resolution and temporal SNR of 3D-EPI examinations with nonselective excitation were optimized at a clinical 3T MR scanner in five healthy subjects using a clinical head/neck coil. A CEST presaturation module for low power relayed nuclear Overhauser enhancement and amide proton transfer contrast was applied as an example. The suggested postprocessing included motion correction, dynamic B0 correction, denoising, and B1 correction and was compared to an established 3D-gradient echo-based sequence. RESULTS: CEST examinations were performed at 1.8 mm nominal isotropic resolution in 4.3 s per presaturation offset. In contrast to slab-selective 3D or multislice approaches, the whole brain was covered. Repeated examinations at three different B1 values took 13 minutes for 58 presaturation offsets with temporal SNR around 75. The resulting CEST effects revealed significant gray and white matter contrast and were of similar quality across the whole brain. Coefficient of variation across three healthy subjects was below 9\textpercent. CONCLUSION: The suggested protocol enables whole brain coverage at 1.8 mm isotropic resolution and fast acquisition of 4.3 s per presaturation offset. For the fitted CEST amplitudes, high reproducibility was proven, increasing the opportunities of quantitative CEST investigations at 3T significantly.}
DOI BibTeX
Miscellaneous Zebrafish exhibit visual attentional pop-out effects in their behavior Stednitz, S., Li, J., Robson, D., Zhaoping, L. 11th European Zebrafish Meeting, 96, 2020
{Sensory inputs drive our interactions with the world, but the raw sensory inputs contain far more data than could be fully processed by our brain, which is constrained by limited resources. To overcome this processing bottleneck, the brain has evolved attentional selectional mechanisms to select only a fraction of the sensory inputs for deeper processing, and for survival this selection is often rapid and exogenously driven (\textquotedblleftbottom-up\textquotedblright) so that it does not require any higher-order (\textquotedbllefttop-down\textquotedblright) cognitive guidance. Such attentional selection for visual inputs has been studied extensively in mammals, particularly in humans. To study the underlying neural mechanism for bottom-up visual selection, the zebrafish (Danio rerio) is perhaps an ideal model system. They rely on visual input for many behaviors early in development, possess evolutionarily conserved brain regions including a large optic tectum associated with attentional selection, and are amenable to live imaging using genetically encoded fluorescent indicators of neuronal activity. However, bottom-up visual attentional selection behavior has not yet been well characterized in zebrafish. A hallmark of the bottom-up visual selection is pop-out, when a uniquely featured visual item in a background of uniformly featured items, e.g., a red dot among green dots, elicits a rapid orienting behavior. Such orienting can be by gaze shifts (often in humans), and/or by turning of head, limb, tentacles, and body. We first demonstrate the pop-out effect in zebrafish using colored dot stimuli, and measure the orienting responses by shifts in the animal\textquoterights body orientation. We characterize the orientation response in relation to the sensory inputs, in particular the colors and spatial configurations of the dots. Our findings suggest conserved mechanisms of visual attention across evolution, and lay the groundwork for imaging studies to describe the underlying neuronal computations.}
BibTeX
Miscellaneous pH mapping of brain tissue by a deep neural network trained on 9.4T CEST MRI data: pH-deepCEST Mueller, S., Glang, F., Scheffler, K., Zaiss, M. 2020 ISMRM & SMRT Virtual Conference & Exhibition, 2020
{The pH value is of major importance for most physiological processes and may change due to altered metabolism in pathologies. In the present work, we exploit the inherent dependency of CEST MR data on pH with a new approach: train neural networks to map voxel-by-voxel from multi-B1+ CEST spectra to pH value. Measurements were performed in homogenate of pig brain tissue at 9.4T ultra high field. Prediction of absolute pH values was possible and predictions were stable against inhomogeneity in B1+. We hope this proof of concept might be a first small step towards high-resolution 3D pH maps in vivo.}
BibTeX
Miscellaneous pulseq-CEST: Towards reproducible CEST experiments using an open source sequence definition standard Herz, K. 8. International Workshop on Chemical Exchange Saturation Transfer Imaging (CEST 2020), 2020 BibTeX
Perceiving Systems Ph.D. Thesis Towards Geometric Understanding of Motion Ranjan, A. University of Tübingen, December 2019
<p> The motion of the world is inherently dependent on the spatial structure of the world and its geometry. Therefore, classical optical flow methods try to model this geometry to solve for the motion. However, recent deep learning methods take a completely different approach. They try to predict optical flow by learning from labelled data. Although deep networks have shown state-of-the-art performance on classification problems in computer vision, they have not been as effective in solving optical flow. The key reason is that deep learning methods do not explicitly model the structure of the world in a neural network, and instead expect the network to learn about the structure from data. We hypothesize that it is difficult for a network to learn about motion without any constraint on the structure of the world. Therefore, we explore several approaches to explicitly model the geometry of the world and its spatial structure in deep neural networks. </p> <p> The spatial structure in images can be captured by representing it at multiple scales. To represent multiple scales of images in deep neural nets, we introduce a Spatial Pyramid Network (SpyNet). Such a network can leverage global information for estimating large motions and local information for estimating small motions. We show that SpyNet significantly improves over previous optical flow networks while also being the smallest and fastest neural network for motion estimation. SPyNet achieves a 97% reduction in model parameters over previous methods and is more accurate. </p> <p> The spatial structure of the world extends to people and their motion. Humans have a very well-defined structure, and this information is useful in estimating optical flow for humans. To leverage this information, we create a synthetic dataset for human optical flow using a statistical human body model and motion capture sequences. We use this dataset to train deep networks and see significant improvement in the ability of the networks to estimate human optical flow. </p> <p> The structure and geometry of the world affects the motion. Therefore, learning about the structure of the scene together with the motion can benefit both problems. To facilitate this, we introduce Competitive Collaboration, where several neural networks are constrained by geometry and can jointly learn about structure and motion in the scene without any labels. To this end, we show that jointly learning single view depth prediction, camera motion, optical flow and motion segmentation using Competitive Collaboration achieves state-of-the-art results among unsupervised approaches. </p> <p> Our findings provide support for our hypothesis that explicit constraints on structure and geometry of the world lead to better methods for motion estimation. </p>
PhD Thesis BibTeX
Micro, Nano, and Molecular Systems Article HPLC of monolayer-protected Gold clusters with baseline separation Knoppe, S., Vogt, P. Analytical Chemistry, 91:1603, December 2019
The properties of ultrasmall metal nanoparticles (ca. 10–200 metal atoms), or monolayer-protected metal clusters (MPCs), drastically depend on their atomic structure. For systematic characterization and application, assessment of their purity is of high importance. Currently, the gold standard for purity control of MPCs is mass spectrometry (MS). Mass spectrometry, however, cannot always detect small impurities; MS of certain clusters, for example, ESI-TOF of Au40(SR)24, is not successful at all. We here present a simple reversed-phase HPLC method for purity control of a series of small alkanethiolate-protected gold clusters. The method allows the detection of small impurities with high sensitivity. Linear correlation between alkyl chain length of Au25(SC_n H_(2n+1))18 clusters (n = 6, 8, 10, 12) and their retention time was noticed.
DOI URL BibTeX
Empirical Inference Conference Paper Optimal experimental design via Bayesian optimization: active causal structure learning for Gaussian process networks von Kügelgen, J., Rubenstein, P. K., Schölkopf, B., Weller, A. NeurIPS 2019 Workshop Do the right thing: machine learning and causal inference for improved decision making, December 2019 (Published) arXiv Poster URL BibTeX
Statistical Learning Theory Conference Paper Foundations of Comparison-Based Hierarchical Clustering Ghoshdastidar, D., Perrot, M., von Luxburg, U. Advances in Neural Information Processing Systems 32 (NIPS 2019), NeurIPS, Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Autonomous Learning Conference Paper Assessing Aesthetics of Generated Abstract Images Using Correlation Structure Khajehabdollahi, S., Martius, G., Levina, A. In Proceedings 2019 IEEE Symposium Series on Computational Intelligence (SSCI), 306-313, IEEE, 2019 IEEE Symposium Series on Computational Intelligence (SSCI), December 2019 Arxiv DOI BibTeX
Empirical Inference Conference Paper Fisher Efficient Inference of Intractable Models Liu, S., Kanamori, T., Jitkrittum, W., Chen, Y. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 8790-8800, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Micro, Nano, and Molecular Systems Modern Magnetic Systems Article Soft-magnetic coatings as possible sensors for magnetic imaging of superconductors Ionescu, A., Simmendinger, J., Bihler, M., Miksch, C., Fischer, P., Soltan, S., Schütz, G., Albrecht, J. Supercond. Sci. and Tech., 33:015002, IOP, December 2019
Magnetic imaging of superconductors typically requires a soft-magnetic material placed on top of the superconductor to probe local magnetic fields. For reasonable results the influence of the magnet onto the superconductor has to be small. Thin YBCO films with soft-magnetic coatings are investigated using SQUID magnetometry. Detailed measurements of the magnetic moment as a function of temperature, magnetic field and time have been performed for different heterostructures. It is found that the modification of the superconducting transport in these heterostructures strongly depends on the magnetic and structural properties of the soft-magnetic material. This effect is especially pronounced for an inhomogeneous coating consisting of ferromagnetic nanoparticles.
DOI URL BibTeX
Theory of Inhomogeneous Condensed Matter Article Using the fluctuation-dissipation theorem for nonconservative forces Asheichyk, K., Krüger, M. Physical Review Research, 1(3):033151, American Physical Society (APS), College Park, Maryland, United States, December 2019 (Published) DOI BibTeX
Empirical Inference Conference Paper A Model to Search for Synthesizable Molecules Bradshaw, J., Paige, B., Kusner, M. J., Segler, M., Hernández-Lobato, J. M. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 7935-7947, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Are Disentangled Representations Helpful for Abstract Visual Reasoning? van Steenkiste, S., Locatello, F., Schmidhuber, J., Bachem, O. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 14222-14235, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Autonomous Learning Conference Paper Control What You Can: Intrinsically Motivated Task-Planning Agent Blaes, S., Vlastelica, M., Zhu, J., Martius, G. In Advances in Neural Information Processing Systems (NeurIPS 2019), 12520-12531, Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Accepted)
We present a novel intrinsically motivated agent that learns how to control the environment in the fastest possible manner by optimizing learning progress. It learns what can be controlled, how to allocate time and attention, and the relations between objects using surprise based motivation. The effectiveness of our method is demonstrated in a synthetic as well as a robotic manipulation environment yielding considerably improved performance and smaller sample complexity. In a nutshell, our work combines several task-level planning agent structures (backtracking search on task graph, probabilistic road-maps, allocation of search efforts) with intrinsic motivation to achieve learning from scratch.
PDF Supplementary material NeurIPS Page Project Page Video Poster BibTeX
Intelligent Control Systems Conference Paper Controlling Heterogeneous Stochastic Growth Processes on Lattices with Limited Resources Haksar, R., Solowjow, F., Trimpe, S., Schwager, M. In Proceedings of the 58th IEEE International Conference on Decision and Control (CDC) , 1315-1322, 58th IEEE International Conference on Decision and Control (CDC), December 2019 (Published) PDF BibTeX
Empirical Inference Probabilistic Numerics Conference Paper Convergence Guarantees for Adaptive Bayesian Quadrature Methods Kanagawa, M., Hennig, P. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 6234-6245, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Movement Generation and Control Article Epstein, Project Maven, and Some Reasons to Think About Where We Get Our Funding [Ethical, Legal, and Societal Issues] Bretl, T., Righetti, L., Madhavan, R. IEEE Robotics & Automation Magazine, 26, December 2019 BibTeX
Empirical Inference Conference Paper Exploiting the modularity of deep networks to generate visual counterfactuals Besserve, M., Mehrjou, A., Sun, R., Schölkopf, B. NeurIPS 2019 - Workshop on Shared Visual Representations in Human & Machine Intelligence, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Flex-Convolution Groh*, F., Wieschollek*, P., Lensch, H. P. A. Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, 11361:105-122, Lecture Notes in Computer Science, (Editors: Jawahar, C. V. and Li, Hongdong and Mori, Greg and Schindler, Konrad), Springer International Publishing, December 2019, *equal contribution (Published) DOI BibTeX
Article Growing the Humanoid Robotics Community Righetti, L., Sugihara, T., Metta, G., Yamane, K. IEEE Robotics & Automation Magazine, 26:136-137, December 2019 (Published) BibTeX
Haptic Intelligence Article Hierarchical Task-Parameterized Learning from Demonstration for Collaborative Object Movement Hu, S., Kuchenbecker, K. J. Applied Bionics and Biomechanics, 2019(9765383):1-25, December 2019 (Published)
Learning from demonstration (LfD) enables a robot to emulate natural human movement instead of merely executing preprogrammed behaviors. This article presents a hierarchical LfD structure of task-parameterized models for object movement tasks, which are ubiquitous in everyday life and could benefit from robotic support. Our approach uses the task-parameterized Gaussian mixture model (TP-GMM) algorithm to encode sets of demonstrations in separate models that each correspond to a different task situation. The robot then maximizes its expected performance in a new situation by either selecting a good existing model or requesting new demonstrations. Compared to a standard implementation that encodes all demonstrations together for all test situations, the proposed approach offers four advantages. First, a simply defined distance function can be used to estimate test performance by calculating the similarity between a test situation and the existing models. Second, the proposed approach can improve generalization, e.g., better satisfying the demonstrated task constraints and speeding up task execution. Third, because the hierarchical structure encodes each demonstrated situation individually, a wider range of task situations can be modeled in the same framework without deteriorating performance. Last, adding or removing demonstrations incurs low computational load, and thus, the robot’s skill library can be built incrementally. We first instantiate the proposed approach in a simulated task to validate these advantages. We then show that the advantages transfer to real hardware for a task where naive participants collaborated with a Willow Garage PR2 robot to move a handheld object. For most tested scenarios, our hierarchical method achieved significantly better task performance and subjective ratings than both a passive model with only gravity compensation and a single TP-GMM encoding all demonstrations.
DOI BibTeX
Empirical Inference Conference Paper Invert to Learn to Invert Putzky, P., Welling, M. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 444-454, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Kernel Stein Tests for Multiple Model Comparison Lim, J. N., Yamada, M., Schölkopf, B., Jitkrittum, W. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 2240-2250, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Probabilistic Numerics Conference Paper Limitations of the empirical Fisher approximation for natural gradient descent Kunstner, F., Hennig, P., Balles, L. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 4158-4169, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Micro, Nano, and Molecular Systems Patent Method and apparatus for fabricating a component Melde, K., Fischer, P. (EP3034281B1), December 2019
The invention relates to a method of fabricating a component having a one-, two- or three-dimensional geometry, wherein the shape of the component is obtained by utilizing an acoustic field and the effect of acoustic forces and by fixating a shape that forms in the acoustic field. This may be achieved by accumulating a material, like e. g. discrete particles in the pressure nodes formed by the acoustic field. Furthermore, the invention relates to an apparatus for fabricating a component, including an acoustic source device for forming an acoustic field that gives rise to a shape, including shapes formed by a particle distribution by acoustic forces, and a fixation device for fixating a shape of the particle distribution. Applications of the invention are available in the fields of fabricating materials with arbitrary shapes, e.g. for rapid prototyping purposes, and the assembly of materials.
BibTeX
Empirical Inference Talk Multivariate coupling estimation between continuous signals and point processes Safavi, S., Logothetis, N., Besserve, M. Neural Information Processing Systems 2019 - Workshop on Learning with Temporal Point Processes, December 2019 (Published) Talk video URL BibTeX
Empirical Inference Conference Paper On the Fairness of Disentangled Representations Locatello, F., Abbati, G., Rainforth, T., Bauer, S., Schölkopf, B., Bachem, O. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 14584-14597, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Optics and Sensing Laboratory Autonomous Motion Conference Paper On the Transfer of Inductive Bias from Simulation to the Real World: a New Disentanglement Dataset Gondal, M. W., Wüthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Schölkopf, B., Bauer, S. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 15714-15725, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Perceiving the arrow of time in autoregressive motion Meding, K., Janzing, D., Schölkopf, B., Wichmann, F. A. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 2303-2314, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Practical and Consistent Estimation of f-Divergences Rubenstein, P. K., Bousquet, O., Djolonga, J., Riquelme, C., Tolstikhin, I. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 4072-4082, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Ph.D. Thesis Real Time Probabilistic Models for Robot Trajectories Gomez-Gonzalez, S. Technical University Darmstadt, Germany, December 2019 (Published) BibTeX
Empirical Inference Ph.D. Thesis Robot Learning for Muscular Systems Büchler, D. Technical University Darmstadt, Germany, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Selecting causal brain features with a single conditional independence test per feature Mastakouri, A., Schölkopf, B., Janzing, D. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 12532-12543, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Empirical Inference Conference Paper Stochastic Frank-Wolfe for Composite Convex Minimization Locatello, F., Yurtsever, A., Fercoq, O., Cevher, V. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 14246-14256, (Editors: H. Wallach and H. Larochelle and A. Beygelzimer and F. d’Alché-Buc and E. Fox and R. Garnett), Curran Associates, Inc., 33rd Annual Conference on Neural Information Processing Systems, December 2019 (Published) URL BibTeX
Micro, Nano, and Molecular Systems Article Acoustic hologram enhanced phased arrays for ultrasonic particle manipulation Cox, L., Melde, K., Croxford, A., Fischer, P., Drinkwater, B. Phys. Rev. Applied, 12:064055, November 2019
The ability to shape ultrasound fields is important for particle manipulation, medical therapeutics and imaging applications. If the amplitude and/or phase is spatially varied across the wavefront then it is possible to project ‘acoustic images’. When attempting to form an arbitrary desired static sound field, acoustic holograms are superior to phased arrays due to their significantly higher phase fidelity. However, they lack the dynamic flexibility of phased arrays. Here, we demonstrate how to combine the high-fidelity advantages of acoustic holograms with the dynamic control of phased arrays in the ultrasonic frequency range. Holograms are used with a 64-element phased array, driven with continuous excitation. Moving the position of the projected hologram via phase delays which steer the output beam is demonstrated experimentally. This allows the creation of a much more tightly focused point than with the phased array alone, whilst still being reconfigurable. It also allows the complex movement at a water-air interface of a “phase surfer” along a phase track or the manipulation of a more arbitrarily shaped particle via amplitude traps. Furthermore, a particle manipulation device with two emitters and a single split hologram is demonstrated that allows the positioning of a “phase surfer” along a 1D axis. This paper opens the door for new applications with complex manipulation of ultrasound whilst minimising the complexity and cost of the apparatus.
DOI URL BibTeX
Haptic Intelligence Conference Paper Deep Neural Network Approach in Electrical Impedance Tomography-Based Real-Time Soft Tactile Sensor Park, H., Lee, H., Park, K., Mo, S., Kim, J. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), (999)7447-7452, IEEE, 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), November 2019 (Published)
Recently, a whole-body tactile sensing have emerged in robotics for safe human-robot interaction. A key issue in the whole-body tactile sensing is ensuring large-area manufacturability and high durability. To fulfill these requirements, a reconstruction method called electrical impedance tomography (EIT) was adopted in large-area tactile sensing. This method maps voltage measurements to conductivity distribution using only a few number of measurement electrodes. A common approach for the mapping is using a linearized model derived from the Maxwell's equation. This linearized model shows fast computation time and moderate robustness against measurement noise but reconstruction accuracy is limited. In this paper, we propose a novel nonlinear EIT algorithm through Deep Neural Network (DNN) approach to improve the reconstruction accuracy of EIT-based tactile sensors. The neural network architecture with rectified linear unit (ReLU) function ensured extremely low computational time (0.002 seconds) and nonlinear network structure which provides superior measurement accuracy. The DNN model was trained with dataset synthesized in simulation environment. To achieve the robustness against measurement noise, the training proceeded with additive Gaussian noise that estimated through actual measurement noise. For real sensor application, the trained DNN model was transferred to a conductive fabric-based soft tactile sensor. For validation, the reconstruction error and noise robustness were mainly compared using conventional linearized model and proposed approach in simulation environment. As a demonstration, the tactile sensor equipped with the trained DNN model is presented for a contact force estimation.
DOI BibTeX
Empirical Inference Conference Paper Improving Local Trajectory Optimisation using Probabilistic Movement Primitives Shyam, R. A., Lightbody, P., Das, G., Liu, P., Gomez-Gonzalez, S., Neumann, G. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2666-2671, IEEE, International Conference on Intelligent Robots and Systems 2019 (IROS) , November 2019 (Published) DOI BibTeX
Empirical Inference Conference Paper Experience Reuse with Probabilistic Movement Primitives Stark, S., Peters, J., Rueckert, E. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 1210-1217, IEEE, November 2019 (Published) DOI BibTeX
Movement Generation and Control Conference Paper Learning to Explore in Motion and Interaction Tasks Bogdanovic, M., Righetti, L. Proceedings 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2686-2692, IEEE, 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), November 2019, ISSN: 2153-0866 (Published)
Model free reinforcement learning suffers from the high sampling complexity inherent to robotic manipulation or locomotion tasks. Most successful approaches typically use random sampling strategies which leads to slow policy convergence. In this paper we present a novel approach for efficient exploration that leverages previously learned tasks. We exploit the fact that the same system is used across many tasks and build a generative model for exploration based on data from previously solved tasks to improve learning new tasks. The approach also enables continuous learning of improved exploration strategies as novel tasks are learned. Extensive simulations on a robot manipulator performing a variety of motion and contact interaction tasks demonstrate the capabilities of the approach. In particular, our experiments suggest that the exploration strategy can more than double learning speed, especially when rewards are sparse. Moreover, the algorithm is robust to task variations and parameter tuning, making it beneficial for complex robotic problems.
DOI BibTeX