Search

Explainable AI provides insights to users into the why for model predictions, offering potential for users to better understand and trust a model, and to recognize and correct AI predictions that are incorrect. Prior research on human and explainable AI interactions has focused on measures such as interpretability, trust, and usability of the explanation. There are mixed findings whether explainable AI can improve actual human decision-making and the ability to identify the problems with the underlying model. Using real datasets, we compare objective human decision accuracy without AI (control), with an AI prediction (no explanation), and AI prediction with explanation. We find providing any kind of AI prediction tends to improve user decision accuracy, but no conclusive evidence that explainable AI has a meaningful impact. Moreover, we observed the strongest predictor for human decision accuracy was AI accuracy and that users were somewhat able to detect when the AI was correct vs. incorrect, but this was not significantly affected by including an explanation. Our results indicate that, at least in some situations, the why information provided in explainable AI may not enhance user decision-making, and further research may be needed to understand how to integrate explainable AI into real systems.

Situation awareness (SA) is a widely used cognitive construct in human factors, often theoretically posited to be a critical causal factor and/or construct for performance. However, there are conce...

Augmented Reality (AR) and Mixed Reality (MR) have been two of the most explosive research topics in the last few years. Head-Mounted Devices (HMDs) are essential intermediums for using AR and MR technology, playing an important role in the research progress in these two areas. Behavioral research with users is one way of evaluating the technical progress and effectiveness of HMDs. In addition, AR and MR technology is dependent upon virtual interactions with the real environment. Thus, conditions in real environments can be a significant factor for AR and MR measurements with users. In this paper, we survey 87 environmental-related HMD papers with measurements from users, spanning over 32 years. We provide a thorough review of AR- and MR-related user experiments with HMDs under different environmental factors. Then, we summarize trends in this literature over time using a new classification method with four environmental factors, the presence or absence of user feedback in behavioral experiments, and ten main categories to subdivide these papers (e.g., domain and method of user assessment). We also categorize characteristics of the behavioral experiments, showing similarities and differences among papers.

We investigated human understanding of different network visualizations in a large-scale online experiment. Three types of network visualizations were examined: node-link and two different sorting variants of matrix representations on a representative social network of either 20 or 50 nodes. Understanding of the network was quantified using task time and accuracy metrics on questions that were derived from an established task taxonomy. The sample size in our experiment was more than an order of magnitude larger (N = 600) than in previous research, leading to high statistical power and thus more precise estimation of detailed effects. Specifically, high statistical power allowed us to consider modern interaction capabilities as part of the evaluated visualizations, and to evaluate overall learning rates as well as ambient (implicit) learning. Findings indicate that participant understanding was best for the node-link visualization, with higher accuracy and faster task times than the two matrix visualizations. Analysis of participant learning indicated a large initial difference in task time between the node-link and matrix visualizations, with matrix performance steadily approaching that of the node-link visualization over the course of the experiment. This research is reproducible as the web-based module and results have been made available at: https://osf.io/qct84/.

The increasing use of Artificial Intelligence (AI) systems in face recognition and video processing in recent times creates higher stakes for their application in daily life. Increasingly, critical decisions are being made using these AI systems in application domains such as employment, finance, and crime prevention. These applications are done through the use of more abstract concepts such as emotions, trait evaluations (e.g., trustworthiness), and behavior (e.g., deception). These abstract concepts are learned by the AI system using the verbal and non-verbal cues from the human subject stimuli (e,g., facial expressions, movements, audio, text) for inference. Because the use of AI systems often happens in high stakes scenarios, it is of utmost importance that the AI system participating in the decision-making process is highly reliable and credible. In this paper, we specifically consider the feasibility of using such an AI system for deception detection. We examine if deception can be caught using multimodal aspects such as facial expressions and movements, audio cues, video cues, etc. We experiment using three different datasets with varying degrees of deception to explore the problem of deception detection. We also study state-of-the-art deception detection systems and investigate whether we can extend their algorithm into new datasets. We conclude that there is a lack of reasonable evidence that AI-based deception detection is generalizable over different scenarios of lying (lying deliberately, lying under duress, and lying through half-truths) and that in the future additional factors will need to be considered to make such a claim.

Successful navigation generally requires an accurate spatial representation of the environment. However, under guided navigation (i.e. route directions are provided by a GPS system or car passenger) no detailed representation of space is necessary because one only needs to follow directions. The first experiment separated components of navigation, spatial decision-making (planning and selection of routes) and navigational control (executing complex actions such as stopping and steering) for environmental learning. Spatial knowledge was assessed using multiple measures (spatial updating using immersive, head-mounted virtual reality and map construction). More accurate environmental knowledge was acquired when spatial decision-making was present at learning, ps < .001, ds > .81. No difference in spatial knowledge was found between active navigation and spatial decision-making by itself, ps > .83, ds < .07. This finding indicates that the weakened spatial knowledge with control alone (guided navigation) can be explained by the absence of spatial decision-making. Experiments 2 and 3 investigated the use of aids for ameliorating the loss of spatial knowledge for guided navigation with auditory directions. In Experiment 2, guided navigation was augmented using visual aids (primarily egocentric) which reinforced the locations of landmarks. In Experiment 3, guided navigation was augmented by adding cardinal directions (e.g. north), which are exocentric, to the auditory guidance, presented either with monaural sound or stereo sound which corresponded to environmental headings. Spatial knowledge in Experiments 2 and 3 was assessed using route replication, novel route execution, and map construction. Contrary to the hypotheses, the aids did not improve all measures of environmental knowledge. Instead, there was a match between the particular reference frame reinforced by the aid and the relevant frame of reference for the measure of spatial knowledge. In Experiment 2, the visual aids improved accuracy for replicating routes, p = .01, d = .75, this measure is egocentric. In Experiment 3, the stereo and monaural cardinal directions improved accuracy for executing a novel route, a directional measure of survey knowledge, ps < .06, ds > .52, which is egocentric and exocentric.

Augmented-reality (AR) displays render virtual content overlaid on the real world. AR offers great potential for Soldiers to immediately perceive relevant tactical information without the need to consult a physical map or other reference material, or divert their eyes away from the immediate environment. However, realizing the potential benefits of using AR requires that the conformal subset of virtual content is perceived in the correct physical location. Conformal virtual content must be correctly associated with the physical location, object, or person intended to be augmented. In this report, we provide emerging recommendations aimed at improving spatial perception of virtual content.

SA is a widely used cognitive construct in human factors, summarized as “knowing what is going on.” Generally, SA is theoretically posited to be a critical causal factor and/or construct for performance. However, some researchers have raised concerns that SA may be circular and also that SA may lack the appropriate psychological mechanisms relevant to performance. We address these conflicting perspectives using meta-analysis to evaluate the specific and general patterns of associations among SA-performance effect sizes. Specifically, we focus on the validity of SA for performance—the degree to which SA represents or captures the relevant psychological processes and mechanisms related to task performance. From the empirical literature, we coded associations of eight unique measures of SA with (task) performance: 492 effects from 38 papers met the systematic review inclusion criteria. In contrast to SA’s broadly theorized fundamental link with performance, the magnitude of most meta-analytic mean effect sizes for SA measures was limited to medium or lower effects. Although there was a significant overall mean effect, its magnitude was also limited (r = 0.24). In addition, there was high unexplained systematic variation with an enormous plausible range for individual effects (r = -0.20 to 0.60). The meta-analytic results are inconsistent with theories postulating SA is fundamental to performance. Instead, SA’s validity for performance tends to be, on average, weak with large variations among effects. Therefore, theories may need to be revised. Furthermore, even presuming SA is causally linked to performance as generally theorized, improvements in SA (such as SA-based design and training) may not correspond to meaningful increases in task performance.

Objective A human factors engineering-based intervention aimed at the modification of task behavior to increase adherence to best practices and the reduction of central line-associated bloodstream infections (CLABSI). The hypothesis was tested that a central line maintenance kit would improve adherence and reduce CLABSI compared with a standard, nonkit-based method of performing central line maintenance. Design The study design was a 29-month prospective, interventional, nonrandomized, observational, and clinical research study using a pre–post implementation assessment. Setting The study was conducted at a tertiary hospital in the southwestern United States, with participants recruited from a total of 6 patient units (including intensive care units and general wards) . Participants A total of 95 nurses and 151 patients volunteered to participate in the study. Intervention A central line maintenance kit was developed that incorporated human factors engineering design principles. This kit was implemented hospitalwide during the clinical study to assess the intervention's influence on protocol adherence and clinical outcomes compared with a preimplementation control condition (no kit use). Results The results of this clinical observations study suggest that a human factors engineering-based kit improved adherence to best practices during central line maintenance. In addition, the number of CLABSIs was significantly reduced during the postimplementation period. Conclusions The application of human factors engineering design principles in the development of medical kits can improve protocol adherence and clinical outcomes.

In light of the Army’s intent to leverage advances in Artificial Intelligence (AI) for augmenting dismounted Soldier Lethality through the development of in-scope and Heads-Up Display (HUD)-based Automatic Target Recognition (ATR) systems, the Combat Capabilities Development Command U. S. Army Research Laboratory’s Human Research and Engineering Directorate (CCDC-ARL/HRED) has identified several critical gaps that must be addressed in order to effectively team the Soldier with ATR for the desired augmented Lethality. One of these areas pertains to the way in which ATR is displayed and requires a thorough understanding and leveraging of relevant cognitive processes that will enable use of this technology. Additionally, insufficient consideration of perceptual, attentional, and cognitive capabilities increases the risk of burdening the Soldier with excessive, unnecessary, or distracting representations of information, which may impede Lethality rather than augment it. HRED’s planned and ongoing research is intended to develop novel mechanisms through which Soldiers teamed with ATR will perform more adaptively and effectively than either the Soldier or the intelligent system could accomplish individually. Based on HRED’s significant expertise in the cognitive sciences, coupled with familiarity with the military-relevant domain spaces, the following initial recommendations for ATR information display requirements are made:1.ATR highlighting should leverage a non-binary display schema to continuously encode threat information (e.g., target class/identity, uncertainty, and prioritization).2.ATR highlighting should be integrated with the target itself instead of functioning as a discrete feature of the display (i.e., highlight the target rather than highlighting a region with the target inside).3.Information about threat certainty or classification confidence (which can also include priority) should be embedded into ATR highlighting.4.Yellow highlights may offer advantages for display5.Changing information (e.g., target certainty) should be accomplished through formation or modification of highlight gradients rather than sudden changes in the display.6.Human performance evaluations of ATR should consider the incorporation of changing threat states and contexts into scenarios for more relevant findings.7.Human performance evaluations of ATR should consider the incorporation of uncued (non-highlighted) targets and miscued targets (false identifications; e.g. ATR identifies non-threat as threat) for more relevant findings.

The Army plans to integrate artificial intelligence (AI)/machine learning (ML) and other intelligent decision-making aids into future dismounted Warfighter systems to augment situational awareness and target acquisition capabilities. However, due to the unique constraints of dismounted operations, successful implementation of intelligent decisionmaking aids in dismounted systems necessitates a human-in-the-loop approach, which includes the ability for the Warfighter to provide feedback to the autonomous system. Human-in-the-loop feedback can augment current machine learning techniques by reducing the size of datasets needed to train algorithms and allow algorithms to be flexible and adaptive to changing battlespace conditions. As such, research is required to define the bidirectional interactions between man and machine in this context, to optimize human-intelligent agent teaming for the dismounted Warfighter. In this paper, we focus on a specific application of dismounted Human-AI interaction to weapon mounted target acquisition (small arms fire control systems) and discuss issues pertaining to an important component of this optimization: how intelligent information is communicated to the end user. We consider how the intelligent information is presented to the Warfighter, and what underlying cognitive and perceptual processes can be leveraged to optimize teamed decision making. Such factors are critical to the successful implementation of human-in-the-loop AI in dismounted applications and ultimately the effectiveness of intelligent decision-making aids.

Technical advances in cyber-attack attribution continues to show incremental improvement. A growing interest in the role of the human in perception management, and decision-making suggest that other aspects of human cognition may be able to help inform attribution, and other aspects of cyber security such as defending and training. Values shape behaviors and cultural values set norms for groups of people. Therefore, they should be considered when modeling behaviors. The lack of studies in this area requires exploration and foundational work to learn the limits of this area of research. This chapter highlights some of the findings of some of the recent studies.

Objective: We investigated how increases in task-relevant information affect human decision-making performance, situation awareness (SA), and trust in a simulated command-and-control (C2) environment. Background: Increased information is often associated with an improvement of SA and decision-making performance in networked organizations. However, previous research suggests that increasing information without considering the task relevance and the presentation can impair performance. Method: We used a simulated C2 task across two experiments. Experiment 1 varied the information volume provided to individual participants and measured the speed and accuracy of decision making for task performance. Experiment 2 varied information volume and information reliability provided to two participants acting in different roles and assessed decision-making performance, SA, and trust between the paired participants. Results: In both experiments, increased task-relevant information volume did not improve task performance. In Experiment 2, increased task-relevant information volume reduced self-reported SA and trust, and incorrect source reliability information led to poorer task performance and SA. Conclusion: These results indicate that increasing the volume of information, even when it is accurate and task relevant, is not necessarily beneficial to decision-making performance. Moreover, it may even be detrimental to SA and trust among team members. Application: Given the high volume of available and shared information and the safety-critical and time-sensitive nature of many decisions, these results have implications for training and system design in C2 domains. To avoid decrements to SA, interpersonal trust, and decision-making performance, information presentation within C2 systems must reflect human cognitive processing limits and capabilities.

In this paper we provide a critical analysis with met- rics that will inform guidelines for designing distributed systems for Collective Situational Understanding (CSU). CSU requires both collective insight—i.e., accurate and deep understanding of a situation derived from uncertain and often sparse data and collective foresight—i.e., the ability to predict what will happen in the future. When it comes to complex scenarios, the need for a distributed CSU naturally emerges, as a single monolithic approach not only is unfeasible: it is also undesirable. We therefore propose a principled, critical analysis of AI techniques that can support specific tasks for CSU to derive guidelines for designing distributed systems for CSU.

Massive amounts of data are being collected, stored, and analyzed for various business and marketing purposes. While such data analysis is critical for many applications, it could also violate the privacy of individuals. This paper describes the issues involved in designing a privacy aware data management framework for collecting, storing, and analyzing the data. We also discuss behavioral aspects of data sharing as well as aspects of a formal framework based on rewriting rules that encompasses the privacy aware data management framework.

In this paper we compare human understanding of information represented in a natural language (NL) to a type of artificial language, called a Controlled Natural Language (CNL). Potential applications for CNLs include decision support and conversational agents, but currently there is limited empirical research on the understandability of CNLs for untrained humans. We investigate a particular type of CNL, called Controlled English (CE), which was designed to be a simplified, artificial subset of natural language that is both human readable and unambiguous for fast and accurate machine processing. We quantify and compare human understanding of NL and CE using accuracy and speed for language statements. The statements described entities (people and objects) and relations (actions) among entities with the ground-truth represented using visual diagrams. Participants responded whether the statements matched the diagram (yes/no). In Experiment I, we found accuracy for NL and CE was comparable, although the speed for understanding CE was slower. To further examine the role of speed, we induced time pressure in Experiment II. We found both the accuracy and speed for CE was lower than NL. These results indicate that if people have sufficient time, understanding for CE can be equivalent to NL. However, with limited time the accuracy and speed for understanding NL is better than CE. Our findings indicate that both accuracy and speed of CNLs should be evaluated. Furthermore, under time pressure there can be meaningful differences in accuracy and speed between different ways of representing information. Understanding for methods of representing machine information has potential implications for situation understanding and management with human-machine interaction and collaboration.

Due to copyright restrictions, the access to the full text of this article is only available via subscription. In this paper we present a methodology to exploit human-machine coalitions for situational understanding. Situational understanding refers to the ability to relate relevant information and form logical conclusions, as well as identify gaps in information. This process for comprehension of the meaning information requires the ability to reason inductively, for which we will exploit the machines’ ability to ‘learn’ from data. However, important phenomena are often rare in occurrence with high degrees of uncertainty, thus severely limiting the availability of instance data for training, and hence the applicability of many machine learning approaches. Therefore, we present the benefits of Subjective Bayesian Networks—i.e., Bayesian Networks with imprecise probabilities—for situational understanding, and the role of conversational interfaces for supporting decision makers in the evolution of situational understanding. Army Research Laboratory ; Ministry of Defence

Peer-to-peer lending is hypothesized to help equalize economic opportunities for the world's poor. We empirically investigate the "flat-world" hypothesis, the idea that globalization eventually leads to economic equality, using crowdfinancing data for over 660,000 loans in 220 nations and territories made between 2005 and 2013. Contrary to the flat-world hypothesis, we find that peer-to-peer lending networks are moving away from flatness. Furthermore, decreasing flatness is strongly associated with multiple variables: relatively stable patterns in the difference in the per capita GDP between borrowing and lending nations, ongoing migration flows from borrowing to lending nations worldwide, and the existence of a tie as a historic colonial. Our regression analysis also indicates a spatial preference in lending for geographically proximal borrowers. To estimate the robustness for these patterns for future changes, we construct a network of borrower and lending nations based on the observed data. Then, to perturb the network, we stochastically simulate policy and event shocks (e.g., erecting walls) or regulatory shocks (e.g., Brexit). The simulations project a drift towards rather than away from flatness. However, levels of flatness persist only for randomly distributed shocks. By contrast, loss of the top borrowing nations produces more flatness, not less, indicating how the welfare of the overall system is tied to a few distinctive and critical country-pair relationships.

—Controlled Natural Language (CNL) has great potential to support human-machine interaction (HMI) because it provides an information representation that is both human readable and machine processable.We investigated the effectiveness of a CNL-based conversational interface for HMI in a behavioural experiment called Simple Human Experiment Regarding Locally Observed Collective Knowledge (SHERLOCK). In SHERLOCK, individuals acted in groups to discover and report information to the machine using natural language (NL), which the machine then processed into CNL. The machine fused responses from different users to form a common operating picture, a dashboard showing the level of agreement for distinct information. To obtain information to add to this dashboard, users explored the real world in a simulated crowd-sourced sensing scenario. This scenario represented a simplified, controlled analogue for tactical intelligence (i.e., direct intelligence of the environment), which is key for rapidly planning military, law enforcement, and emergency operations. Overall, despite close to zero training, 74% of the users inputted NL that was machine interpretable and addressed the assigned tasks. An experimental manipulation aimed to increase user-machine interaction, however, did not improve performance as hypothesised. Nevertheless, results indicate the conversational interface may be effective in assisting humans with collection and fusion of information in a crowd-sourcing context.

Summary Stereotypes can harm human performance, especially when activated in individuals with diminished working memory capacity (WMC). Performance implications for the stereotype of poor driving in older adults were investigated. Using a sample of older adults, WMC (the ability to maintain task goals and ignore distractions) and driving performance [brake reaction time (RT), following distance, and crashes] were assessed, the latter using a high-fidelity simulator. Elderly participants under stereotype threat with reduced WMC exhibited slower brake RTs and longer following distances compared with a control condition that was not threatened. This driving profile was characteristic of cognitive distraction. Stereotype threat has clear consequences for human performance in a common real-world task—driving—that is critical to public safety. Furthermore, these findings suggest caution in how the media and public policy communicate information about older adult driving.Copyright © 2015 John Wiley & Sons, Ltd.

The spread of opinions in social networks is dependent on structural properties of the network and the individual characteristics of its nodes. To capture this dependence, several abstract models of such spread were proposed. First, we model the difference between the dynamics of opinion spread in communities with a static social network versus a dynamic social network. Here we use the theoretical model of spread of opinions called the Binary Agreement Model based on the naming game. Using this model, we study potential mechanisms for the dependencies observed in the data by matching model generated evolution of opinion with the empirically observed evolution in the data. Second, we examine the unique set of behavioral network data (based on electronic logs of dyadic contact via smartphones) collected at the University of Notre Dame. The participants are a sample of members of the entering class of freshmen in the fall of 2011 whose opinions on a wide variety of political and social issues have been regularly recorded—at the beginning and end of each semester—for the last three years. Using this data set, we measure the evolution of participants’ opinions and ascertain how much this evolution depends on the cultural traits of individuals and the structural properties of social networks that they form. Our analysis of our empirical dataset shows that ties among people who are more likely to share opinions (e.g. same race, gender, or socioeconomic class) decay at a slower rate than ties among persons who are likely to have different opinions. The analysis also indicates that the partner selection of individuals is associated sharing a (political) opinion. These results offer an assessment of the level of impact of culture and social network dynamics on the evolution of opinions in multi-cultural social networks.

Most risk analysis models systematically underestimate the probability and impact of catastrophic events (e.g., economic crises, natural disasters, and terrorism) by not taking into account interconnectivity and interdependence of risks. To address this weakness, we propose the Cascading Alternating Renewal Process (CARP) to forecast interconnected global risks. However, assessments of the model's prediction precision are limited by lack of sufficient ground truth data. Here, we establish prediction precision as a function of input data size by using alternative long ground truth data generated by simulations of the CARP model with known parameters. We illustrate the approach on a model of fires in artificial cities assembled from basic city blocks with diverse housing. The results confirm that parameter recovery variance exhibits power law decay as a function of the length of available ground truth data. Using CARP, we also demonstrate estimation using a disparate dataset that also has dependencies: real-world prediction precision for the global risk model based on the World Economic Forum Global Risk Report. We conclude that the CARP model is an efficient method for predicting catastrophic cascading events with potential applications to emerging local and global interconnected risks., 12 pages 6 figures