Publications (new)

This work proposes SafeDQN, which allows making the behavior of autonomous vehicles safe and interpretable while still being efficient. SafeDQN offers an understandable, semantic trade-off between the expected risk and the utility of actions while being algorithmically transparent. We show that SafeDQN finds interpretable and safe driving policies for various scenarios and demonstrate how state-of-the-art saliency techniques can help assess risk and utility.

We propose IALE, an imitation learning scheme that imitates the selection of the best-performing expert heuristic at each stage of the learning cycle in a batch-mode pool-based setting. We use Dagger to train a transferable policy on a dataset and later apply it to different datasets and deep classifier architectures. The policy reflects on the best choices from multiple expert heuristics given the current state of the active learning process, and learns to select samples in a complementary way that unifies the expert strategies. Our experiments on well-known image datasets show that we outperform state of the art imitation learners and heuristics.

This work proposes ‘incremental data-uploading’, a novel encoding pattern for high dimensional data that tackles these challenges. We spread the encoding gates for the feature vector of a given data point throughout the quantum circuit with parameterized gates in between them. This encoding pattern results in a better representation of data in the quantum circuit with a minimal pre-processing requirement. We show the efficiency of our encoding pattern on a classification task using the MNIST and Fashion-MNIST datasets, and compare different encoding methods via classification accuracy and the effective dimension of the model.

We propose a novel neural network architecture for multivariate time series that notably improves the classification and trajectory regression performance in online handwriting recognition. We achieve this by jointly learning the cross-entropy loss in combination with distance and similarity losses.

We propose a transfer learning (TL) method that exploits realistic synthetic Channel State Information (CSI) obtained with the Quasi Deterministic Radio channel Generator (QuaDRiGa), used to pre-train the CNN-based fingerprint model so that it can be adapted to any real (NLoS) propagation scenario with a low number of real training samples. Our experiments show that the positioning accuracy using fine-tuning improves by 37% in changed and by 19% in new environments.

In this paper, we focus on models for online handwriting recognition, one particular type of spatio-temporal data. The data is observed from a sensor-enhanced pen with the goal to classify written characters. We conduct a broad evaluation of aleatoric (data) and epistemic (model) UQ based on two prominent techniques for Bayesian inference, Stochastic Weight Averaging-Gaussian (SWAG) and Deep Ensembles. Next to a better understanding of the model, UQ techniques can detect out-of-distribution data and domain shifts when combining right-handed and left-handed writers (an underrepresented group).

We show that VQ-DQN approaches are subject to instabilities that cause the learned policy to diverge, study the extent to which this afflicts reproducibility of established results based on classical simulation, and perform systematic experiments to identify potential explanations for the observed instabilities. Additionally, we execute RL algorithms on an actual quantum processing unit (an IBM Quantum Device) and investigate differences in behaviour between simulated and physical quantum systems that suffer from implementation deficiencies. Our experiments show that it cannot be conclusively decided if known quantum approaches, even if simulated without physical imperfections, can provide an advantage as compared to classical approaches.

We evaluate multiple neural network architectures containing encoding (compressing) and decoding (reconstructing) components and compare them to the state-of-the-art compression techniques Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT). We show that fully connected autoencoders achieve the lowest error, outperforming both DCT and DWT. Further experiments prove that the reconstructed CSI can be used for positioning with only mild performance deterioration at a compression of >97% and even when trained on a different environment.

We introduce a novel pipeline that builds on recent advances in imitation learning and that can generate safe and efficient behavior policies. We combine an RL step that solves for safe behavior through the introduction of safety distances with a subsequent innovative safe extraction of decision tree policies. The resulting decision tree is not only easy to interpret, it is also safer than the neural network policy trained for safety. Additionally, we formally prove the safety of trained RL agents for linearized system dynamics, showing that the learned and extracted policy successfully avoids all catastrophic events.

We examined the Google/Apple Exposure Notification Framework (ENF), which in combination with the front-end, i.e., the German Corona-Warn-App (CWA), enables digital CT in Germany. With centimeter accurate optical reference systems we show that optimal parameters are application-specific. However, they cause impractical high resource costs. In contrast, optimized general parameters offer an everyday compromise between energy costs, applicability, accuracy, and reliability of the ENF.

We present a novel representation learning approach using Siamese Metric Learning that approximates a distance preserving low-dimensional representation and that learns to estimate reasonable solutions to the assignment problem. To this end, we employ a Temporal Convolutional Network architecture that we extend with a gating mechanism to enable learning from sparse data, leading to solutions to the assignment problem exhibiting varying degrees of sparsity.

We propose to use generative deep learning models for out-of-distribution detection based on channel impulse responses (CIRs). We use a Variational Autoencoder (VAE) to predict an anomaly score for the channel of a TOF-based Ultra-wideband (UWB) system. Our experiments show that a VAE trained only on line-of-sight (LOS) training data generalizes well to new environments and detects non-line-of-sight CIRs with an accuracy of 85%. We also show that integrating our anomaly score into a TOF-based extended Kalman filter (EKF) improves tracking performance by over 25%.

The 2020 IPIN Competition included 5 diverse off-site off-site Tracks, each resembling real use cases and challenges for indoor positioning. The results in terms of participation and accuracy of the proposed systems have been encouraging. The best performing competitors obtained a third quartile of error of 1 m for the Smartphone Track and 0.5 m for the Foot-mounted IMU Track. While not running on physical systems, but only as algorithms, these results represent impressive achievements.

We propose a 1D Convolutional Neural Network to identify optimal FDPoAs as ToA directly from the raw CSI. We train our model with realistic 5G channel model data synthesized by QuaDRiGa. Metrics such as Delay Spread (DS), k-Factor (kF), and SNR are appropriate to cover most LoS-NLoS constellations. Our approach consistently outperforms the state of the art by about 17% for SNRs below -10 dB.

We develop a Machine Learning (ML) classifier that predicts quality metrics from a 9-degrees of freedom inertial measurement unit, while we simultaneously guide and track the work processes with a HoloLens AR system. In our user study, 6 test subjects perform typical assembly tasks with our system. We evaluate the tracking accuracy of the system based on a precise optical reference system and evaluate the classification of each work step quality based on the collected ground truth data. Our evaluation shows a tracking accuracy of fast dynamic movements of 4.92mm and our classifier predicts the actions carried out with mean F1 value of 93.8% on average.

We introduce a process monitoring system that uses inertial, magnetic field and audio sensors that we attach as add-ons to hand-held tools. The sensor data is analyzed via embedded classification algorithms and our system directly provides feedback to workers during the execution of work processes. We outline the special requirements caused by vastly different tools and show how to automatically train and deploy new ML models.

We use domain randomization to study simulation-to-reality transfer for a high-speed collision avoidance scenario. We train the policy not only on a single version of the setup but on several variations. Our experiments show that the resulting policy is able to generalize much better to different values for the vehicle speed and distance from the obstacle compared to policies trained in the non-randomized version of the setup.

Video: https://download.cmutschler.de/publications/2020/ITSC/0703_VP.mp4
Slides: https://download.cmutschler.de/publications/2020/ITSC/0703_PS.pdf

We study the applicability of today’s popular AR systems (Apple ARKit, Google ARCore, and Microsoft Hololens) in an industrial context (large area of 1,600m2, long walking distances of 60m between cubicles, and dynamic environments with volatile natural features). Out of the box, these AR systems are far from useful even for normal motion behavior. They accumulate an average error of about 17m per 120m, with a scaling error of up to 14.4cm/m that is quasi-directly proportional to the path length. By adding natural features, the tracking reliability can be improved, but not enough.

We apply different convolutional neural network architectures (ResNet, Encoder, FCN) to detect non line-of-sight (NLOS) channel conditions directly from the CIR raw data. A realistic measurement campaign is used to train and evaluate the algorithms. The proposed methods highly outperform the featurebased ML baselines while still using low network complexities. We also show that the models generalize well to unknown receivers and environments and that positioning filters benefit significantly from the identification of NLOS measurements.

We propose an approach to reconstruct gait motions from a single head-mounted accelerometer. We train our models to map head motions to corresponding ground truth gait phases. To reconstruct leg motion, the models predict gait phases to trigger equivalent synthetic animations. We designed four models: a threshold-based, a correlation-based, a Support Vector Machine (SVM) -based and a bidirectional long-term short-term memory (BLSTM) -based model. Our experiments show that, while the BLSTM approach is the most accurate, only the correlation approach runs on a mobile VR system in real time with sufficient accuracy. Our user study with 21 test subjects examined the effects of our approach on simulator sickness and showed significantly less negative effects on disorientation.

We propose a hybrid filter that combines a convolutional neural network (CNN) and a bidirectional recurrent neural network (BLSTM) (that extract spatial features from the sensor signals and track their temporal relationships) with a linear Kalman filter (LKF) that improves the velocity estimates. Our experiments show the robustness against different movement states and changes in orientation, even in highly dynamic situations. It also generalizes well to different and varying movement speeds and provides accurate and precise velocity estimates.

We propose a novel online handwriting dataset acquired from 119 writers consisting of 31,275 uppercase and lowercase English alphabet character recordings (52 classes) as part of the UbiComp 2020 Time Series Classification Challenge. Our novel OnHW-chars dataset allows for the evaluations of uppercase, lowercase and combined classification tasks, on both writer-dependent (WD) and writer-independent (WI) classes and we show that properly tuned machine learning pipelines as well as deep learning classifiers (such as CNNs, LSTMs, and BiLSTMs) yield accuracies up to 90 % for the WD task and 83 % for the WI task for uppercase characters.

We propose ViPR, a novel modular architecture for longterm 6DoF VO that leverages temporal information and synergies between absolute pose estimates (from PoseNet-like modules) and relative pose estimates (from FlowNet-based modules) by combining both through recurrent layers. Experiments on known datasets and on our own Industry dataset show that our modular design outperforms state of the art in long-term navigation tasks.

We ML/DL to estimate a human velocity. Our approach is robust to varying motion states and orientation changes in dynamic situations. On data from a single un-calibrated IMU, our novel recurrent model not only outperforms the state-of-the-art on instantaneous velocity (≤0.10 m/s) and on traveled distance (≤29 m/km). It can also generalize to different and varying rates of motion and provides accurate and precise velocity estimates.