The central thesis of this paper is that the technology of intelligent, autonomous machines gives rise to novel fault modes that are not seen in other types of automation. As a consequence, autonomous systems provide new vectors for cyber-attack with the potential consequence of subversion, degraded behavior or outright failure of the autonomous system. While we can only pursue the analogy so far, maladaptive behavior and the other symptoms of these fault modes in some cases may resemble those found in humans. The term “psychopathology” is applied to fault modes of the human mind, but as yet we have no equivalent area of study for intelligent, autonomous machines. This area requires further study in order to document and explain the symptoms of unique faults in intelligent systems, whether they occur in nominal conditions or as a result of an outside, purposeful attack. By analyzing algorithms, architectures and what can go wrong with autonomous machines, we may a) gain insight into mechanisms of intelligence; b) learn how to design out, work around or otherwise mitigate these new failure modes; c) identify potential new cyber-security risks; d) increase the trustworthiness of machine intelligence. Vigilance and attention management mechanisms are identified as specific areas of risk.

Communication is social by nature, and reveals psycho-social dimensions about an actor’s perceptions of themself and others. While grammar and spell-check can help polish the presentation of communication, it does not reflect the way that a message will be received in a particular social space. A means to analyze the communication for actor beliefs can help the author and others understand the underlying social climate and message that is being transmitted. NASA has identified the need to monitor individual behavioral health and team dynamics as crucial to ensuring high performance and mission success. We describe an application that integrates theories from sociolinguistics with natural language processing techniques to successfully detect individual moods, attitudes, and team dynamics relevant to long duration exploration class missions. The methods were used to analyze data gathered from human subject experiments at three diverse analog studies, with results showing high correlation with subject self-reports and third party observations. We discuss preliminary results and implications for the tool’s potential wide-spread use.

Toward constructing a multimodal conversation agentsystem which can be used to interview elderly patients with dementia, we propose a turn taking mechanism based on recognition of the subjects attitude as to whether the subject has (or relinquish) the right to speak. A key strategy in the recognition task is to extract features from pausing behavior in subject’s spontaneous speech and to fuse multimodal signals (gaze, head motion, and speech). In this paper, we focus on evaluation of the recognition module used in guiding turn taking. To evaluate it, we collect multimodal data corpus from 42 dyadic conversations between subjects with dementia and the virtual agent we have developed as a prototype and annotate subject’s multimodal data manually. In experiments, we validate recognition models trained multimodal dataset by machine learning methods.Experimental results shows that pause features are effective to improve the attitude recognition accuracy and the accuracy is improved up to 88%.

High-level representations of uncertainty, such as probabilistic logics and programs, have been around for decades. Lifted inference was initially motivated by the need to make reasoning algorithms high-level as well. While the lifted inference community focused on machine learning applications, the high-level reasoning goal has received less attention recently. We revisit the idea and look at the capabilities of the latest techniques in lifted inference. This lets us conclude that lifted inference is strictly more powerful than propositional inference on high-level reasoning tasks.

Robots autonomy has been widely focused on in the newspapers with a trend towards anthropomorphism that is likely to mislead people and conceal or disguise the technical reality. This paper aims at reviewing the different technical aspects of robots autonomy. First we propose a definition allowing to distinguish robots from devices that are not robots. Then autonomy is defined and considered as a relative notion within a framework of authority sharing between the decision functions of the robot and the human being. Several technical issues are mentioned according to three points of view: (i) the robot, (ii) the human operator and (iii) the interaction between the operator and the robot. Some key questions that should be carefully dealt with for future robotic systems are given at the end of the paper.

While today's structured knowledge bases (e.g., Freebase) contain a sizable collection of information about entities, from celebrities and locations to concepts and common objects, there is a class of knowledge that has minimal coverage: actions. A large-scale knowledge base of actions would provide an opportunity for computinng devices to aid and support people's reasoning about their own actions and outcomes, leading to improved decision-making and goal achievement. In this short paper, we describe our first efforts towards building a distributional representation of actions and their outcomes, as learned from the timelines of individuals posting experiential microblogs.

The representation of dates and their relationship to time and duration has long been recognized as an important problem in commonsense reasoning. However, existing date ontologies, such as OWL-Time and Date-Time Foundation Vocabulary from the Object Modeling Group, take either over-simplistic or convoluted approaches to defining the key semantics for dates. We show that such approaches are inadequate and provide an improved solution: a first-order Date Ontology that is an extension of the Process Specification Language and an existing duration ontology. Rather than treat dates as a class of timepoints, we axiomatize dates as a class of complex activities which have multiple periodic occurrences. We consider two modules of the Date Ontology, and characterize the models of the Date Ontology up to elementary equivalence.

In this paper, we study the effects of different ways of producing greetings in human-robot interaction. We first generated computer utterances of verbal greetings, whose intonation contours, we then manipulated using Praat. Each utterance was matched with a video of a robot waving a greeting at the observer. Altogether, the experiment uses two lexical items ( hello vs. hi ), three robots and four different intonation contours. The videos were distributed over different questionnaires so that each participant only got to see each robot once. The results reveal that native speakers of English rate the robots significantly different concerning friendliness, assertiveness, and engagement depending on the intonation contours. However, these effects differ for the different lexical items, and the apparently non-conventional hi with rising intonation contour was in fact rated as most engaging.

Traditional approaches to knowledge base completion have been based on symbolic representations. Low-dimensional vector embedding models proposed recently for this task are attractive since they generalize to possibly unlimited sets of relations. A significant draw- back of previous embedding models for KB completion is that they merely support reasoning on individual relations (e.g., bornIn ( X, Y ) ⇒ nationality ( X, Y ) ). In this work, we develop models for KB completion that support chains of reasoning on paths of any length using compositional vector space models. We construct compositional vector representations for the paths in the KB graph from the semantic vector representations of the binary relations in that path and perform inference directly in the vector space. Unlike previous methods, our approach can generalize to paths that are unseen in training and, in a zero-shot setting, predict target relations without supervised training data for that relation.

Commonsense reasoning proves to be an essential tool for natural-language access to data. In a deductive approach to this problem, language processing technology translates English queries into a first-order logical form, which is regarded as a conjecture to be established by a theorem prover. Subject domain knowledge is encoded in an axiomatic theory equipped with links to appropriate databases. Commonsense reasoning is necessary to disambiguate the query, to connect the query with relevant tables in the databases, to deal with logical relationships in the query, and to achieve interoperability between disparate databases. This is illustrated with examples from a proof-of-concept system called Quest, which deals with queries over business enterprise data for an industrial QA system.