Open Access

In robot-assisted therapy for individuals with Autism Spectrum Disorder, the workload of therapists during a therapeutic session is increased if they have to control the robot manually. To allow therapists to focus on the interaction with the person instead, the robot should be more autonomous, namely it should be able to interpret the person's state and continuously adapt its actions according to their behaviour. In this paper, we develop a personalised robot behaviour model that can be used in the robot decision-making process during an activity; this behaviour model is trained with the help of a user model that has been learned from real interaction data. We use Q-learning for this task, such that the results demonstrate that the policy requires about 10,000 iterations to converge. We thus investigate policy transfer for improving the convergence speed; we show that this is a feasible solution, but an inappropriate initial policy can lead to a suboptimal final return.

The dataset contains the following data from successful and failed executions of the Toyota HSR robot placing a book on a shelf. RGB images from the robot's head camera Depth images from the robot's head camera Rendered images of the robot's 3D model from the point of view of the robot's head camera Force-torque readings from a wrist-mounted force-torque sensor Joint efforts, velocities and positions extrinsic and intrinsic camera calibration parameters frame-level anomaly annotations The anomalies that occur during execution include: the manipulated book falling down books on the shelf being disturbed significantly camera occlusions robot being disturbed by an external collision The dataset is split into a train, validation and test set with the following number of trials: Train: 48 successful trials Validation: 6 successful trials Test: 60 anomalous trials and 7 successful trials

Short summary This dataset accompanies our paper A. Mitrevski, P. G. Plöger, and G. Lakemeyer, "Representation and Experience-Based Learning of Explainable Models for Robot Action Execution," in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. Contents There are three zip archives included, each of them a dump of a MongoDB database corresponding to one of the three experiments in the paper: Grasping a drawer handle (handle_drawer_logs.zip) Grasping a fridge handle (handle_fridge_logs.zip) Pulling an object (pull_logs.zip) All three experiments were performed with a Toyota HSR. Only the data necessary for learning the models used in our experiments are included here. Usage After unzipping the archives, each database can be restored with the command mongorestore [directory_name] This will create a MongoDB database with the name of the directory (handle_drawer_logs, handle_fridge_logs, and pull_logs). Code for processing the data and model learning can be found in our <a href="https://github.com/alex-mitrevski/explainable-robot-execution-models">GitHub repository.

The accurate forecasting of solar radiation plays an important role for predictive control applications for energy systems with a high share of photovoltaic (PV) energy. Especially off-grid microgrid applications using predictive control applications can benefit from forecasts with a high temporal resolution to address sudden fluctuations of PV-power. However, cloud formation processes and movements are subject to ongoing research. For now-casting applications, all-sky-imagers (ASI) are used to offer an appropriate forecasting for aforementioned application. Recent research aims to achieve these forecasts via deep learning approaches, either as an image segmentation task to generate a DNI forecast through a cloud vectoring approach to translate the DNI to a GHI with ground-based measurement (Fabel et al., 2022; Nouri et al., 2021), or as an end-to-end regression task to generate a GHI forecast directly from the images (Paletta et al., 2021; Yang et al., 2021). While end-to-end regression might be the more attractive approach for off-grid scenarios, literature reports increased performance compared to smart-persistence but do not show satisfactory forecasting patterns (Paletta et al., 2021). This work takes a step back and investigates the possibility to translate ASI-images to current GHI to deploy the neural network as a feature extractor. An ImageNet pre-trained deep learning model is used to achieve such translation on an openly available dataset by the University of California San Diego (Pedro et al., 2019). The images and measurements were collected in Folsom, California. Results show that the neural network can successfully translate ASI-images to GHI for a variety of cloud situations without the need of any external variables. Extending the neural network to a forecasting task also shows promising forecasting patterns, which shows that the neural network extracts both temporal and momentarily features within the images to generate GHI forecasts.

Risk-based authentication (RBA) aims to protect users against attacks involving stolen passwords. RBA monitors features during login, and requests re-authentication when feature values widely differ from previously observed ones. It is recommended by various national security organizations, and users perceive it more usable and equally secure than equivalent two-factor authentication. Despite that, RBA is still only used by very few online services. Reasons for this include a lack of validated open resources on RBA properties, implementation, and configuration. This effectively hinders the RBA research, development, and adoption progress. To close this gap, we provide the first long-term RBA analysis on a real-world large-scale online service. We collected feature data of 3.3 million users and 31.3 million login attempts over more than one year. Based on the data, we provide (i) studies on RBA’s real-world characteristics, and its configurations and enhancements to balance usability, security, and privacy, (ii) a machine learning based RBA parameter optimization method to support administrators finding an optimal configuration for their own use case scenario, (iii) an evaluation of the round-trip time feature’s potential to replace the IP address for enhanced user privacy, and (iv) a synthesized RBA data set to reproduce this research and to foster future RBA research. Our results provide insights on selecting an optimized RBA configuration so that users profit from RBA after just a few logins. The open data set enables researchers to study, test, and improve RBA for widespread deployment in the wild.

We introduce canonical weight normalization for convolutional neural networks. Inspired by the canonical tensor decomposition, we express the weight tensors in so-called canonical networks as scaled sums of outer vector products. In particular, we train network weights in the decomposed form, where scale weights are optimized separately for each mode. Additionally, similarly to weight normalization, we include a global scaling parameter. We study the initialization of the canonical form by running the power method and by drawing randomly from Gaussian or uniform distributions. Our results indicate that we can replace the power method with cheaper initializations drawn from standard distributions. The canonical re-parametrization leads to competitive normalization performance on the MNIST, CIFAR10, and SVHN data sets. Moreover, the formulation simplifies network compression. Once training has converged, the canonical form allows convenient model-compression by truncating the parameter sums.

Deep learning has become a one-size-fits-all solution for technical and business domains thanks to its flexibility and adaptability. It is implemented using opaque models, which unfortunately undermines the outcome trustworthiness. In order to have a better understanding of the behavior of a system, particularly one driven by time series, a look inside a deep learning model so-called posthoc eXplainable Artificial Intelligence (XAI) approaches, is important. There are two major types of XAI for time series data, namely model-agnostic and model-specific. Model-specific approach is considered in this work. While other approaches employ either Class Activation Mapping (CAM) or Attention Mechanism, we merge the two strategies into a single system, simply called the Temporally Weighted Spatiotemporal Explainable Neural Network for Multivariate Time Series (TSEM). TSEM combines the capabilities of RNN and CNN models in such a way that RNN hidden units are employed as attention weights for the CNN feature maps temporal axis. The result shows that TSEM outperforms XCM. It is similar to STAM in terms of accuracy, while also satisfying a number of interpretability criteria, including causality, fidelity, and spatiotemporality.

As cameras are ubiquitous in autonomous systems, object detection is a crucial task. Object detectors are widely used in applications such as autonomous driving, healthcare, and robotics. Given an image, an object detector outputs both the bounding box coordinates as well as classification probabilities for each object detected. The state-of-the-art detectors are treated as black boxes due to their highly non-linear internal computations. Even with unprecedented advancements in detector performance, the inability to explain how their outputs are generated limits their use in safety-critical applications in particular. It is therefore crucial to explain the reason behind each detector decision in order to gain user trust, enhance detector performance, and analyze their failure. Previous work fails to explain as well as evaluate both bounding box and classification decisions individually for various detectors. Moreover, no tools explain each detector decision, evaluate the explanations, and also identify the reasons for detector failures. This restricts the flexibility to analyze detectors. The main contribution presented here is an open-source Detector Explanation Toolkit (DExT). It is used to explain the detector decisions, evaluate the explanations, and analyze detector errors. The detector decisions are explained visually by highlighting the image pixels that most influence a particular decision. The toolkit implements the proposed approach to generate a holistic explanation for all detector decisions using certain gradient-based explanation methods. To the author’s knowledge, this is the first work to conduct extensive qualitative and novel quantitative evaluations of different explanation methods across various detectors. The qualitative evaluation incorporates a visual analysis of the explanations carried out by the author as well as a human-centric evaluation. The human-centric evaluation includes a user study to understand user trust in the explanations generated across various explanation methods for different detectors. Four multi-object visualization methods are provided to merge the explanations of multiple objects detected in an image as well as the corresponding detector outputs in a single image. Finally, DExT implements the procedure to analyze detector failures using the formulated approach. The visual analysis illustrates that the ability to explain a model is more dependent on the model itself than the actual ability of the explanation method. In addition, the explanations are affected by the object explained, the decision explained, detector architecture, training data labels, and model parameters. The results of the quantitative evaluation show that the Single Shot MultiBox Detector (SSD) is more faithfully explained compared to other detectors regardless of the explanation methods. In addition, a single explanation method cannot generate more faithful explanations than other methods for both the bounding box and the classification decision across different detectors. Both the quantitative and human-centric evaluations identify that SmoothGrad with Guided Backpropagation (GBP) provides more trustworthy explanations among selected methods across all detectors. Finally, a convex polygon-based multi-object visualization method provides more human-understandable visualization than other methods. The author expects that DExT will motivate practitioners to evaluate object detectors from the interpretability perspective by explaining both bounding box and classification decisions.

While most approaches individually exploit unstructured data from the biomedical literature or structured data from biomedical knowledge graphs, their union can better exploit the advantages of such approaches, ultimately improving representations of biology. Using multimodal transformers for such purposes can improve performance on context dependent classication tasks, as demonstrated by our previous model, the Sophisticated Transformer Trained on Biomedical Text and Knowledge Graphs (STonKGs). In this work, we introduce ProtSTonKGs, a transformer aimed at learning all-encompassing representations of protein-protein interactions. ProtSTonKGs presents an extension to our previous work by adding textual protein descriptions and amino acid sequences (i.e., structural information) to the text- and knowledge graph-based input sequence used in STonKGs. We benchmark ProtSTonKGs against STonKGs, resulting in improved F1 scores by up to 0.066 (i.e., from 0.204 to 0.270) in several tasks such as predicting protein interactions in several contexts. Our work demonstrates how multimodal transformers can be used to integrate heterogeneous sources of information, paving the foundation for future approaches that use multiple modalities for biomedical applications.

Original dataset Tasks: Two self-motion perception tasks (Move-to-Target and Adjust-Target) and a Size/Distance Perception task Conditions: in-pool / in-lab and supine/upright Self-reported vection experience