Search

Explanations are generated to accompany a model decision indicating features of the input data that were the most relevant towards the model decision. Explanations are important not only for understanding the decisions of deep neural network, which in spite of their their huge success in multiple domains operate largely as abstract black boxes, but also for other model classes such as gradient boosted decision trees. In this work, we propose methods, using both Bayesian and Non-Bayesian approaches to augment explanations with uncertainty scores. We believe that uncertainty augmented saliency maps can help in better calibration of the trust between human analyst and the machine learning models.

In elite sports, there is an opportunity to take advantage of rich and detailed datasets generated across multiple threads of the sporting business. Challenges currently exist due to time constraints to analyse the data, as well as the quantity and variety of data available to assess. Artificial Intelligence (AI) techniques can be a valuable asset in assisting decision makers in tackling such challenges, but deep AI skills are generally not held by those with rich experience in sporting domains. Here, we describe how certain commonly available AI services can be used to provide analytic assistance to sports experts in exploring, and gaining insights from, typical data sources. In particular, we focus on the use of Natural Language Processing and Conversational Interfaces to provide users with an intuitive and time‐saving toolkit to explore their datasets and the conclusions arising from analytics performed on them. We show the benefit of presenting powerful AI and analytic techniques to domain experts, showing the potential for impact not only at the elite level of sports, where AI and analytic capabilities may be more available, but also at a more grass‐roots level where there is generally little access to specialist resources. The work described in this paper was trialled with Leatherhead Football Club, a semi‐professional team that, at the time, were based in the English 7th tier of football.

Traditional service-based applications, in fixed networks, are typically constructed and managed centrally and assume stable service endpoints and adequate network connectivity. Constructing and maintaining such applications in dynamic heterogeneous wireless networked environments, where limited bandwidth and transient connectivity are commonplace, presents significant challenges and makes centralized application construction and management impossible. In this demonstration we present an architecture which is capable of providing an adaptable and resilient method for on-demand decentralized construction and management of complex time-critical applications in such environments. The approach uses a Vector Symbolic Architecture (VSA) to compactly represent an application as a single semantic vector that encodes the service interfaces, workflow, and the time-critical constraints required. By extending existing services interfaces, with a simple cognitive layer that can interpret and exchange the vectors, we show how the required services can be dynamically discovered and interconnected in a completely decentralized manner. There are a large number of workflow systems designed to work in various scientific domains, including support for the Internet of Things (IoT). One such workflow system is Node-RED, which is designed to bring workflow-based programming to IoT. The main focus of this demonstration is to show how we can migrate Node-RED workflows into a decentralized execution environment, so that such workflows can run on Edge networks.

As neural networks gain importance with several successful applications of them, this paper raises the question of how they can be applied in the context of coalition operations. A key challenge in military coalition operations is that of energy and severe bandwidth constraints. We address this challenge by exploring the use of Deep Neural Networks (DNNs) and splitting them across multiple edge nodes. Further, we explore the idea of using spiking neural networks that can lower the energy consumption significantly. Preliminary results show that both these approaches can have significant impact on coalition operations.

Future military coalition operations will increasingly rely on machine learning (ML) methods to improve situational awareness. The coalition context presents unique challenges for ML: the tactical environment creates significant computing and communications limitations while also having to deal with an adversarial presence. Further, coalition operations must operate in a distributed manner, while coping with the constraints posed by the operational environment. Envisioned ML deployments in military assets must be resilient to these challenges. Here, we focus on the susceptibility of ML models to be poisoned (during training) or fooled (after training) by adversarial inputs. We review recent work on distributed adversarial ML, and present new results from our own investigations into model poisoning attacks on distributed learning systems without a central parameter aggregation node.

Situational understanding is impossible without causal reasoning and reasoning under and about uncertainty, i.e. probabilistic reasoning and reasoning about the confidence in the uncertainty assessment. We therefore consider the case of subjective (uncertain) Bayesian networks. In previous work we notice that when observations are out of the ordinary, confidence decreases because the relevant training data, effective instantiations, to determine the probabilities for unobserved variables, on the basis of the observed variables, is significantly smaller than the size of the training data, the total number of instantiations. It is therefore of primary importance for the ultimate goal of situational understanding to be able to efficiently determine the reasoning paths that lead to low confidence whenever and wherever it occurs: this can guide specific data collection exercises to reduce such an uncertainty. We propose three methods to this end, and we evaluate them on the basis of a case-study developed in collaboration with professional intelligence analysts.

Recent advances in Machine Learning (ML) have profoundly changed many detection, classification, recognition and inference tasks. Given the complexity of the battlespace, ML has the potential to revolutionise how Coalition Situation Understanding is synthesised and revised. However, many issues must be overcome before its widespread adoption. In this paper we consider two - interpretability and adversarial attacks. Interpretability is needed because military decision-makers must be able to justify their decisions. Adversarial attacks arise because many ML algorithms are very sensitive to certain kinds of input perturbations. In this paper, we argue that these two issues are conceptually linked, and insights in one can provide insights in the other. We illustrate these ideas with relevant examples from the literature and our own experiments.

Deep neural networks have achieved near-human accuracy levels in various types of classification and prediction tasks including images, text, speech, and video data. However, the networks continue to be treated mostly as black-box function approximators, mapping a given input to a classification output. The next step in this human-machine evolutionary process — incorporating these networks into mission critical processes such as medical diagnosis, planning and control — requires a level of trust association with the machine output. Typically, statistical metrics are used to quantify the uncertainty of an output. However, the notion of trust also depends on the visibility that a human has into the working of the machine. In other words, the neural network should provide human-understandable justifications for its output leading to insights about the inner workings. We call such models as interpretable deep networks. Interpretability is not a monolithic notion. In fact, the subjectivity of an interpretation, due to different levels of human understanding, implies that there must be a multitude of dimensions that together constitute interpretability. In addition, the interpretation itself can be provided either in terms of the low-level network parameters, or in terms of input features used by the model. In this paper, we outline some of the dimensions that are useful for model interpretability, and categorize prior work along those dimensions. In the process, we perform a gap analysis of what needs to be done to improve model interpretability.

Our particular research in the Distributed Analytics\ud and Information Science International Technology Alliance\ud (DAIS ITA) is focused on ”Anticipatory Situational Understanding\ud for Coalitions”.\ud This paper takes the concrete example of detecting and\ud predicting traffic congestion in the UK road transport network\ud from existing generic sensing sources, such as real-time CCTV\ud imagery and video, which are publicly available for this purpose.\ud This scenario has been chosen carefully as we believe that in\ud a typical city, all data relevant to transport network congestion\ud information is not generally available from a single unified source,\ud and that different organizations in the city (e.g. the weather office,\ud the police force, the general public, etc.) have their own different\ud sensors which can provide information potentially relevant to\ud the traffic congestion problem. In this paper we are looking at\ud the problem of (a) identifying congestion using cameras that,\ud for example, the police department may have access to, and (b)\ud fusing that with other data from other agencies in order to (c)\ud augment any base data provided by the official transportation\ud department feeds. By taking this coalition approach this requires\ud using standard cameras to do different supplementary tasks like\ud car counting, and in this paper we examine how well those tasks\ud can be done with RNN/CNN, and other distributed machine\ud learning processes.\ud In this paper we provide details of an initial four-layer\ud architecture and potential tooling to enable rapid formation of\ud human/machine hybrid teams in this setting, with a focus on\ud opportunistic and distributed processing of the data at the edge\ud of the network. In future work we plan to integrate additional\ud data-sources to further augment the core imagery data.

In the wake of rapid maturing of Internet of Things (IoT) approaches and technologies in the commercial sector,\ud the IoT is increasingly seen as a key ‘disruptive’ technology in military environments. Future operational environments\ud are expected to be characterized by a lower proportion of human participants and a higher proportion of\ud autonomous and semi-autonomous devices. This view is reflected in both US ‘third offset’ and UK ‘information\ud age’ thinking and is likely to have a profound effect on how multinational coalition operations are conducted\ud in the future. Much of the initial consideration of IoT adoption in the military domain has rightly focused\ud on security concerns, reflecting similar cautions in the early era of electronic commerce. As IoT approaches\ud mature, this initial technical focus is likely to shift to considerations of interactivity and policy. In this paper,\ud rather than considering the broader range of IoT applications in the military context, we focus on roles for IoT\ud concepts and devices in future intelligence, surveillance and reconnaissance (ISR) tasks, drawing on experience\ud in sensor-mission resourcing and human-computer collaboration (HCC) for ISR. We highlight the importance of\ud low training overheads in the adoption of IoT approaches, and the need to balance proactivity and interactivity\ud (push vs pull modes). As with sensing systems over the last decade, we emphasize that, to be valuable in ISR\ud tasks, IoT devices will need a degree of mission-awareness in addition to an ability to self-manage their limited\ud resources (power, memory, bandwidth, computation, etc). In coalition operations, the management and potential\ud sharing of IoT devices and systems among partners (e.g., in cross-coalition tactical-edge ISR teams) becomes a\ud key issue due heterogeneous factors such as language, policy, procedure and doctrine. Finally, we briefly outline a\ud platform that we have developed in order to experiment with human-IoT teaming on ISR tasks, in both physical\ud and virtual settings.

The rapid identification for drugs-of-abuse in airports is of critical importance. In this study we demonstrate the viability of Raman spectroscopy for the rapid identification of illicit substances in their containers in an airport environment. Raman spectra of drugs-of-abuse in situ were collected using portable Raman spectrometers; this technique offers distinct advantages to government agencies, first responders and forensic scientists working in the security field. We have demonstrated that the spectrometers are able to collect the spectra of suspect powders, including cocaine HCl and d-amphetamine sulphate with unknown constituents rapidly and with a high degree of discrimination. Copyright © 2008 John Wiley & Sons, Ltd.

Reducing poverty is important for those affected, for society and the economy. Poverty remains entrenched in the UK, despite considerable research efforts to understand its causes and possible solutions. The Joseph Rowntree Foundation, with the Centre for Science and Policy at the University of Cambridge, ran a democratic, transparent, consensual exercise involving 45 participants from government, non-governmental organisations, academia and research to identify 100 important research questions that, if answered, would help to reduce or prevent poverty. The list includes questions across a number of important themes, including attitudes, education, family, employment, heath, wellbeing, inclusion, markets, housing, taxes, inequality and power.

While the physics of local field potential (LFP) generation are well-established, the complexity of neural network dynamics means that interpreting LFP measurements in terms of the underlying neural activity is very difficult [1]. Recent studies have therefore investigated forward models of the LFP: calculating the LFP due to known arrangements of neuronal current sources [1]. We use this approach to study the spatio-temporal features of the LFP recorded during persistent gamma-oscillations in vitro. The mechanism by which this activity regime is generated is well-known [2], and many multi-electrode array recordings are available that allow us to compare the spatio-temporal properties of the experimentally measured and simulated LFPs. For simplicity, we implemented a model of a neocortical slice containing only layer 2/3, using our VERTEX simulation tool. The model contained a gamma oscillation-generating region surrounded by regions that did not participate in the gamma oscillation (i.e. the neurons in these regions fired randomly). These enforced conditions allowed us to study the spatial profile of the LFP from differently sized gamma oscillation-generating patches, while minimising the number of variables we changed between simulations. Additionally, we could make observations about properties of the LFP during random activity in the network. We describe three key results from our simulations. Firstly, we found that perisomatic synaptic currents on pyramidal neurons resulting from basket interneuron firing dominate the LFP during gamma oscillations, in agreement with recent experimental results [3,4]. We predict that basket interneurons will also contribute the majority of the LFP signal during random, uncorrelated activity because of the location of their synapses at pyramidal neuron somas. This contribution is amplified during oscillatory activity because of increased basket cell synchrony relative to other neuron types. Secondly, we investigated how gamma-frequency power from a localised gammagenerating region in the model spread due to volume conduction. We found that the spatial spread of the oscillation in the LFP increased above and below the neurons’ somadepth, and depended on the level of the random surround activity. Finally, we investigated frequency scaling in the LFP power spectrum across space. We showed that LFPs at the level of the pyramidal cells’ apical dendrites exhibit a low-pass filtering effect (as predicted previously [5]), which is absent from LFPs recorded at the soma level because of the relative dominance and localisation of the basket cell input. We confirmed that this matched the frequency scaling in our comparison experimental data recorded in macaque neocortical slices, and, based on the frequency characteristics in our model, propose that the experimentally recorded LFP could contain contributions from spiking activity at frequencies down to, and perhaps below, 100 Hz.

Coalition operations environments are characterised\ud by the need to share intelligence, surveillance and reconnaissance\ud services. Increasingly, such services are based on artificial intelligence\ud (AI) and machine learning (ML) technologies. Two key\ud issues in the exploitation of AI/ML services are robustness and\ud interpretability. Employing a diverse portfolio of services can\ud make a system robust to ‘unknown unknowns’. Interpretability\ud — the need for services to offer explanation facilities to engender\ud user trust — can be addressed by a variety of methods to\ud generate either transparent or post hoc explanations according to\ud users’ requirements. This paper shows how a service-provisioning\ud framework for coalition operations can be extended to address\ud specific requirements for robustness and interpretability, allowing\ud automatic selection of service bundles for intelligence, surveillance\ud and reconnaissance tasks. The approach is demonstrated\ud in a case study on traffic monitoring featuring a diverse set of\ud AI/ML services based on deep neural networks and heuristic\ud reasoning approaches.