Driver intention recognition (DIR) studies increasingly rely on deep neural networks. Deep neural networks have achieved top performance for many different tasks. However, apart from image classifications and semantic segmentation for mobile phones, it is not a common practice for components of advanced driver assistance systems to explicitly analyze the complexity and performance of the network’s architecture. Therefore, this paper applies neural architecture search to investigate the effects of the deep neural network architecture on a real-world safety critical application with limited computational capabilities. We explore a pre-defined search space for three deep neural network layer types that are capable to handle sequential data (a long-short term memory, temporal convolution, and a time-series transformer layer), and the influence of different data fusion strategies on the driver intention recognition performance. A set of eight search strategies are evaluated for two driver intention recognition datasets. For the two datasets, we observed that there is no search strategy clearly sampling better deep neural network architectures. However, performing an architecture search improves the model performance compared to the original manually designed networks. Furthermore, we observe no relation between increased model complexity and better driver intention recognition performance. The result indicate that multiple architectures can yield similar performance, regardless of the deep neural network layer type or fusion strategy. However, the optimal complexity, layer type and fusion remain unknown upfront.

Making breast cancer prognosis from gene expression profiles of the primary tumor has become a promising application of deep learning. Yet, to be relevant to real world applications in the clinic and for knowledge discovery, these models must be robust to common distribution shifts. In this study, we evaluate recently proposed methods for improving domain and subgroup shifts. We test the in-distribution and out-of-distribution generalization of multiple episode learning, stochastic weight averaging, group distributionally robust optimization, and a subsampling scheme on one training and four external breast cancer prognosis datasets. The evaluation found that the methods can, to various degrees, improve generalization across domains, although there remain, partially high, generalization gaps. Additionally, in-distribution and out-of-distribution generalization differs between clinical subtypes of breast cancer. Thus, we conclude that further research into methods specifically addressing challenges in breast cancer prognosis from gene expression data are warranted.

Traffic fatalities remain among the leading death causes worldwide. To reduce this figure, car safety is listed as one of the most important factors. To actively support human drivers, it is essential for advanced driving assistance systems to be able to recognize the driver’s actions and intentions. Prior studies have demonstrated various approaches to recognize driving actions and intentions based on in-cabin and external video footage. Given the performance of self-supervised video pre-trained (SSVP) Video Masked Autoencoders (VMAEs) on multiple action recognition datasets, we evaluate the performance of SSVP VMAEs on the Honda Research Institute Driving Dataset for driver action recognition (DAR) and on the Brain4Cars dataset for driver intention recognition (DIR). Besides the performance, the application of an artificial intelligence system in a safety-critical environment must be capable to express when it is uncertain about the produced results. Therefore, we also analyze uncertainty estimations produced by a Bayes-by-Backprop last-layer (BBB-LL) and Monte-Carlo (MC) dropout variants of an VMAE. Our experiments show that an VMAE achieves a higher overall performance for both offline DAR and end-to-end DIR compared to the state-of-the-art. The analysis of the BBB-LL and MC dropout models show higher uncertainty estimates for incorrectly classified test instances compared to correctly predicted test instances.

Detection of anomalous trajectories is an important problem in the surveillance domain. Various algorithms based on learning of normal trajectory patterns have been proposed for this problem. Yet, these algorithms typically suffer from one or more limitations: They are not designed for sequential analysis of incomplete trajectories or online learning based on an incrementally updated training set. Moreover, they typically involve tuning of many parameters, including ad-hoc anomaly thresholds, and may therefore suffer from overfitting and poorly-calibrated alarm rates. In this article, we propose and investigate the Sequential Hausdorff Nearest-Neighbor Conformal Anomaly Detector (SHNN-CAD) for online learning and sequential anomaly detection in trajectories. This is a parameter-light algorithm that offers a well-founded approach to the calibration of the anomaly threshold. The discords algorithm, originally proposed by Keogh et al. , is another parameter-light anomaly detection algorithm that has previously been shown to have good classification performance on a wide range of time-series datasets, including trajectory data. We implement and investigate the performance of SHNN-CAD and the discords algorithm on four different labeled trajectory datasets. The results show that SHNN-CAD achieves competitive classification performance with minimum parameter tuning during unsupervised online learning and sequential anomaly detection in trajectories.

To investigate the impact of visualizing car uncertainty on drivers' trust during an automated driving scenario, a simulator study was conducted. A between-group design experiment with 59 Swedish drivers was carried out where a continuous representation of the uncertainty of the car's ability to autonomously drive during snow conditions was displayed to one of the groups, whereas omitted for the control group. The results show that, on average, the group of drivers who were provided with the uncertainty representation took control of the car faster when needed, while they were, at the same time, the ones who spent more time looking at other things than on the road ahead. Thus, drivers provided with the uncertainty information could, to a higher degree, perform tasks other than driving without compromising with driving safety. The analysis of trust shows that the participants who were provided with the uncertainty information trusted the automated system less than those who did not receive such information, which indicates a more proper trust calibration than in the control group.

Modern healthcare's need for knowledge sharing and bridging the research–practice gap requires new forms of collaboration, in which clinicians of varying clinical and research expertise work together over geographical and organisational borders. To support such distributed communities of practice (CoPs), an understanding of their collaboration processes, outcomes, challenges and enablers is needed. The article examines these issues through a case study of a long-running CoP, the Swedish Oral Medicine Network (SOMNet). SOMNet's main form of collaboration is monthly telephone conference meetings centred on case consultations. Cases are submitted by the clinicians via a Web-based system. The methods used were interviews, observations, and a questionnaire. The work adds to previous research by studying a distributed CoP explicitly focused on supporting the transfer of scientific results from researchers to practitioners. We found that the regular meetings give a rhythm to the community. The centrality of cases means an immediate benefit for the submitter while the community is provided an authentic context for learning. SOMNet yields opportunities for help and learning for diverse expertise levels; the type of benefits is affected by the participant's degree of oral medicine knowledge and collaboration involvement. There are challenges in accommodating varying levels of expertise and encouraging those less experienced to participate. Enablers of the collaboration include the participation of experts, meeting facilitators and well-adapted ITs.

Despite the growing number of systems providing visual analytic support for investigative analysis, few empirical studies include investigations on the analytical reasoning process that needs to be supported. In this paper, we present an approach to evaluate the ability of certain visual representations from an integrated visual-computational environment to support the completion of representative tasks. The problem area studied is the detection and identification of anomalous vessels and situations while monitoring maritime traffic data. This paper presents: (1) a brief review of current evaluation methodologies within information visualization and visual analytics, (2) an analysis of operator's analytical reasoning process (derived from field work in maritime control centers and a literature review on analytical reasoning theories), (3) a list of representative tasks for usability evaluation and (4) an approach to evaluate the use of normal behavioral models representations during the detection process.

The surveillance of large sea areas often generates huge amounts of multidimensional data. Exploring, analyzing and finding anomalous behavior within this data is a complex task. Confident decisions upon the abnormality of a particular vessel behavior require a certain level of situation awareness that may be difficult to achieve when the operator is overloaded by the available information. Based on a visual analytics process model, we present a novel system that supports the acquisition of situation awareness and the involvement of the user in the anomaly detection process using two layers of interactive visualizations. The system uses an interactive data mining module that supports the insertion of the user's knowledge and experience in the creation, validation and continuous update of the normal model of the environment.

Prediction of the enemy's intention is a main issue of threat analysis, and, hence, will be an important part of the C2-systems of tomorrow. A technique that can be useful for this kind of predictions is Bayesian networks (BNs). We have developed a BN for prediction of the enemy's tactical intention, and the implemented BN has been integrated into a ground target simulation framework. The general problem of how to find appropriate prior distributions for BNs has been addressed by developing a tool for data collection, which may make it easier to come up with appropriate prior distributions, by learning conditional probability tables from collected cases, i.e. parameter learning