2

Network and Information Sciences

The Panel on Information Science at the Army Research Laboratory (ARL) conducted its review of selected research and development (R&D) projects of the ARL network and information sciences research core competency at Adelphi, Maryland, on June 18-20, 2019. The research areas reviewed were information sciences and networks and cyber.

ARL research in information sciences seeks to develop new technologies that allow for information acquisition, analysis, reasoning and decision making, and assurance of information and knowledge. This research effort targets technological developments and enhancements that allow for efficient management of information in a dense data environment, extracting knowledge to deploy distributed intelligent systems, and to build effective systems for Multi-Domain Operations (MDO) that will define the battlespace of the future. Technologies resulting from these efforts will have a direct impact on future Army capabilities in C4ISR,¹ networks, intelligent systems, and cybersecurity.

The research projects under the banner of information sciences (IS) had a significant emphasis on artificial intelligence (AI) and machine learning (ML) as applied to diverse areas of Army relevance. Application areas included image understanding, automated language processing, augmented and virtual reality (AR/VR), and learning for control. There was a focus on MDO, and research projects were both foundational and disruptive, while maintaining Army relevance. There is a broad embrace of the transformative potential of AI and ML, especially in the context of how groups of people and autonomous systems can seamlessly collaborate, how technology can further aid with time-sensitive decision making in the presence of massive and diverse sources of information, and how virtual and augmented reality can be integrated in Army operations.

Research in the area of networks and cyber includes projects that fall into the general area of human-robot/machine interactions, with the two principal threads being scene narrative generation for humans by robots and robot learning from human demonstrations. This body of work is ambitious and has the potential to disrupt the way human-robot interactions are considered for future battlefields. Another important research methodology, especially for security issues, is to study systems that have both defense and offense techniques. The “adversarial learning” project is a great example of researchers working on new research areas, with the goal of leveraging what they learn to build better defenses. ML rightly pervades multiple research topics in this thrust.

___________________

¹ C4ISR stands for command, control, communications, computers, intelligence, surveillance, and reconnaissance.

INFORMATION SCIENCES

Accomplishments and Advancements

The research portfolio in information sciences addresses the challenge of overcoming the computational and network resource constraints that affect the efficiency of mission execution, to develop approaches that facilitate and enhance human-machine interaction, and to build human-like capabilities in autonomous systems to identify saliency in visual images. The research portfolio was a mix of basic and applied research, and some of the projects show clear potential for transition into Army applications and products. The researchers are to be complimented for a range of projects that represent a wide swath of Army-relevant IS topics, with particular emphasis on tactical mission, perception and learning, and humans and robots.

Mission success in tactical environments requires the successful deployment and use of edge devices. Here, edge devices refer to devices that employ a distributed computing paradigm that brings computation and data storage closer to the location where it is needed, to improve response times and save bandwidth. The processing and analysis of data from such devices requires significant computational resources that need to be either off-loaded or accessed in a distributed computing network. Both approaches represent vulnerabilities to attack. ARL seeks a secure adaptive computing framework for the tactical environment and is exploring both traditional security approaches and AI-enabled systems that focus on both efficiency and security.

The ARL focus on human-machine interactions stems from an understanding that current “autonomous” systems require continual human involvement. The tactical environment is unstructured and continually changing. As such, it is difficult to provide reliable a priori information about the environment in which the autonomous system will operate. An important focus of ARL research, therefore, is on the human-in-the-loop ML techniques to help develop autonomous agents and systems.

ARL research is also examining techniques for detecting and prioritizing objects and locations within a visible image to enable better autonomous maneuver and operations within a multidomain environment. Toward this end, research is being directed to extend models of visual salience from standard dynamic range (SDR) to high dynamic range (HDR) environments.² The focus is on developing computational models to predict behavioral and neurophysiological effects of HDR luminance.

The research review consisted of a number of presentations on topics such as learning from imitation, risk-aware learning, HDR saliency, and reinforcement learning for adaptable agents. There were also a number of poster presentations in these subject areas, including research on imitation from observation, semantic-based autonomous navigation, learned control policy, and information agents for value assessment. Posters also presented research on saliency in HDR environments, in camouflage, and for moving objects and observers. Research in multitask learning as applied to accent language learning and in morphologically complex language processing was also presented in a poster.

Augmented Reality for Human-Robot Teaming in Field Environments

This project addresses the problem of collaboration between humans and a mobile robot in exploring an unknown environment. Through bidirectional communication, the human and robot know each other’s

___________________

² SDR displays typically use 8 bits of color depth. This limits the contrast ratio for display to about 1,200 to 1. HDR displays typically require a minimum of 10 bits of color depth. In this scheme, it is fairly easy to obtain contrast ratios well over 200,000 to 1. In HDR displays, the backlight is often divided into smaller zones and controlled individually. Matching the brightness of these zones to the overlaid pixel data extends the contrast range for the user.

locations and plans, and can jointly create a map of the shared environment (without global positioning systems [GPS]). Augmented reality allows the human to easily “see” parts of the environment that have already been explored, and to make recommendations on where to explore next. The approach is being validated through experiments conducted at the Military Operations in Urban Terrain (MOUT) Laboratory. The project involves multiple university partners and collaborators. Overall, this is an important area of research for the Army.

Time-Delayed Neural Networks for Acoustic Models Using Multitask Learning as a Domain Adaptation

Global reach and engagement are at the core of Army operations. There is a clear need for improved speech-to-speech translation technologies, especially in languages for which there is limited availability of training data. Research in this project is focused on developing a deep neural network-based multitask learning approach to leverage existing large training sets from another language. Through the use of shared intermediate representation to simultaneously learn speech-to-speech models for both languages, the approach provides for improved accuracy of the models learned. This project represents a prime example of leveraging a state-of-the-art approach to solve a critical need and of significantly enhancing Army operations.

The Salience of Highly Informative but Nonobvious Regions of Interest

Saliency models are designed to focus on obvious features in an image. This research project is directed at developing models that are trained to identify subtle characteristics in an image and to prioritize obvious and nonobvious regions. The research seeks to incorporate semantic information and the embedded relational clues in the model training process. This project is just at the inception stage. The overall aim of the project is important and represents a potentially niche area of Army interest.

Modeling Visual Salience in High Dynamic Range Environments

Visual salience is an important characteristic that is central to the study of visual search. It is used to identify and prioritize images that define a visual scene. Many extant measures of visual saliency are adequate for images with 8-bit SDR. However, these measures need be readdressed in HDR images, where existing theories of luminance perception are less applicable. The project focuses on the development of computational models of visual salience that address shortcomings of standard saliency models. The focus of the work is on modeling interactions of complex features to improve the predictive capability of computational models. In particular, models drawn from neurophysiological studies on HDR are considered promising in the work. Additional model development, training, and testing are required to assess the impact of the work.

Modeling the Salience of Partially Obscured and Camouflaged Objects Under Relative Motion

The research is focused on developing and training spatiotemporal saliency models that improve detection of partially obscured objects. The concept of motion parallax is deployed for the detection of occluded and camouflaged objects, using a spatiotemporal saliency model. Such an approach can contribute to improved situational understanding. Virtual reality simulations were used to generate video data with motion parallax to train deep learning neural networks that will be tested against spatiotemporal test data. The project is a multiyear effort and is at a preliminary stage. While substantial results are yet to

be obtained, preliminary findings indicate that including motion parallax in detection improves the current spatiotemporal saliency models. Additional work is required to assess how camera motion affects obscured object detection.

Learning Control Policies for Perception in Large Environments

This project is directed at addressing challenges related to developing situation awareness for diverse Army missions. The focus of the work is to automatically learn the camera pan, tilt, and zoom parameters for different contexts and scenarios. A deep reinforcement learning-based approach is pursued, and some depth estimation results have been obtained. However, this work is still at a preliminary stage and significantly more tasks are planned to overcome multiple challenges that the project expects to address.

Semantic-Based Autonomous Navigation

The work in this project is motivated by the need to adapt a vehicle’s autonomous driving software to operate effectively in new settings. In an off-line mode, an inverse optimal control approach is used to develop a reward function that visualizes feature maps and trajectory exemplars. This reward function is then used in an on-line mode where features of the operating environment are presented to develop the navigation path. Experimental testing was done at an MOUT site. The project is an excellent example of technology translation from basic research developed through a Collaborative Technology Alliance (CTA) to ARL. The work is clearly significant for Army applications, and related publications are strong.

Information Agent for Assessing Value

This research project work explores modeling of information prioritization that subject matter experts make. The impact of the work can be potentially significant as it can both augment and amplify the human effort that underlies this task. Although some interesting preliminary results have been generated, it would be important to clearly enunciate the rationale for the chosen approach, what competing strategies are possible, and whether the proposed approach performs better than existing solutions to this problem.

Neural Network Models for Low-Resource and Morphologically Complex Language Processing

This project is designed to improve the information extraction capabilities for morphologically complex languages such as Russian and Ukrainian by developing deep neural network-based morphological classifiers for such languages. Neural approaches have not always proven effective in low-data scenarios, and segmental recurrent neural networks and sequence-to-sequence neural networks were deployed for morphological analysis and generation, respectively, with 90-92 percent accuracy. Although these results are promising, it not obvious how significant the improvements are over existing approaches and why commercial (translation or morphological analysis) solutions are not sufficient.

Imitation from Observation

This research focuses on human-in-the-loop ML techniques to develop autonomous AI agents for Army applications. The work represents an extremely productive and substantive collaboration between researchers at ARL and at the University of Texas, Austin, focused on building autonomous systems

through reinforcement learning from human demonstrations of desired behaviors. The research examines strategies to extend behavior cloning and inverse reinforcement learning to the information from observation environment. The work is of high scientific quality, addressing emerging questions in ML and AI—such as how to learn when you do not have large quantities of data. It is also being published in top-tier publications venues for work in this area.

Challenges and Opportunities

The research projects were focused on important areas and researchers were generally able to articulate Army relevance and needs. The overall quality of the research was high, but not uniformly so. In most cases, the posters and presentations provided sufficient details about assumptions made and the major steps followed in the conduct of the research. This was not always the case, and research quality, impact, and understanding would be improved by additional attention to these descriptions in the posters and oral presentations. Some of the presentations did not recognize connections to existing research outside ARL, and in some cases, it was not clear why existing techniques were considered to be inadequate to address the problem. In a few cases, the research problem statements could have been more ambitious.

The research related to imitation from observation was outstanding and potentially disruptive. The work on semantic-based autonomous navigation was a good example of application of technology invented in a CTA that led to a research project at ARL. The project with the goal to learn control policies is clearly highly relevant but can benefit from a widening of scope to consider alternative approaches. Additionally, some interaction with researchers working in saliency may have beneficial impact.

The virtual reality demonstration illustrated a technology of potentially high impact. The project on augmented reality for human-robot teaming represents a promising area for ARL researchers; partnering with the virtual reality research team may offer new avenues for exploration in this context. In particular, research on human-robot teaming could benefit from collaboration with the ongoing work on immersive virtual reality. Doing so would allow for the exploration of immersive approaches to presenting multimodal data to humans.

The research on natural language processing was promising, and work related to application of multitask learning to small-corpus accented speech recognition is clearly an important area for research. The analysis of morphologically complex languages was creative and may fill a need—competing approaches could be considered in this context.

The ongoing work related to information agents for assessing value would benefit from asking more ambitious questions in richer problems, and to bring in other ML methods. Significant effort is directed at analyzing saliency in video; this work is both important and impactful. This work could benefit from consideration of related approaches in the literature, crystallization of goals, and clear plans to achieve them.

NETWORKS AND CYBER

Accomplishments and Advancements

The research portfolio in the area of networks and cyber seeks greater understanding of the general area of diverse communications modalities, and human-robot/machine interactions. It is also focused on a study of adversarial systems as they relate to cybersecurity. ML and AI are central to many of the ongoing research projects.

The research reviewed was high quality and comparable to research conducted at other federal laboratories and educational institutions. For the majority of the research projects, it was clear that the scope of the research was informed by the needs and unique challenges of the Army. Research efforts

were adequately referenced and, in many instances, results from ongoing research were benchmarked against those available in the literature. The research findings are being published in high-quality venues.

In the area of diverse communication modalities, the focus is on developing mobile networking with favorable propagation characteristics in obstructed environments, avoiding detection, and minimizing risk of signal interception, and in general, enhancing network capacity and capabilities. A significant focus is on using mathematical constructs of geometry and topology to explain network characteristics.

In the context of cybersecurity, an important thrust is the security of Army vehicle platforms. Military vehicles contain multiple embedded networked systems that also seek information from external networks for successful operations. In this context, ARL is pursuing research to develop an active defense network for electronically controlled automated vehicle systems. Research in game theory, adversarial ML, and human-machine teaming is being considered in this context.

Research is also directed at developing strategies to ensure robust operation of a “network of things” that have the propensity to fail, exhibit degraded performance, or may be compromised in operation. Approaches based on blockchain technology are being explored in this context. Principles of “fog computing” are also being adopted to address key problem areas.

Some research projects were considered to be exceptionally strong and offer significant potential to contribute to U.S. Army capabilities. Research related to active defense and dynamic watermarking for cyber defense of vehicles and other cyber-physical systems was one such effort where commercial vendors are unlikely to provide solutions. This significant cyber vulnerability represents a pervasive problem in most major Army vehicles.

Another noteworthy project was in the area of narrative generation as it relates to human-robot interaction. The focus of this effort was in developing an understanding and explanation of visual scenery by extracting information and adding captions to video stills that describes relevant scene information and ultimately what transpired. This result alone, if successful, could provide the warfighter a significant workload reduction in processing full motion video or still frame imagery. While not the focus of the ongoing work, this research could potentially also save communications bandwidth by the transmittal of the textual descriptions of the scenes instead of the full motion video.

Low-Power, Low-Frequency Mobile Networking

The goals and objectives of the project were to understand diverse communications modalities for more robust and covert operations, understand physical layer challenges and limitations, improve low probability of detection and low probability of intercept, and exploit autonomous agents that enhance networking capabilities and control radio radiation signatures.

High-frequency communications have favorable propagation characteristics in obstructive terrain but also require physically large antennas. The research leveraged previous very high frequency (VHF) antenna design work on three-dimensional (3D) structural antennas that were cited as having an electrical length of 1/50 lambda, which is physically quite small. The non-Foster design technique used enabled an antenna at VHF to exhibit outstanding electrical length of less than 0.1 meter; the antenna does, however, suffer significantly in transmission efficiency, which is predicted by the fundamental limit. The approach taken by the researcher was to further expand the capabilities of the passive antenna by adding active and passive inductive loading to the antenna, resulting in an increased transmission efficiency. The simulated active proposed design appears to offer larger bandwidth support and a higher transmission efficiency over the passive design.

Multiuser Networking with Improved Low Probability of Detection and Low Probability of Intercept

This research seeks to develop an understanding of the propagation channel by analyzing indoor propagation in an office-type environment at VHF frequencies. An environment of concrete walls with metallic reinforcement at a center frequency of 40 MHz and a bandwidth of 4 MHz was considered in the work. The antennas were composed of dipoles and the environment was simulated using the industry standard finite-difference time-domain approach for electromagnetic simulation and prediction. The research also included a waveform delay-tolerant chip encoding method. The results showed that encoding can tolerate delays of several chips in multipath propagation without severe deterioration in waveform detection performance. The simulation also demonstrated that two users could operate in the channel simultaneously with comparably small chip delays.

Adversarial Machine Learning of Cyber Defense

This research project is directed at studying adversarial ML in the context of network security. The canonical problems center on adversaries injecting data into the network to confuse friendly classifiers. Similar problems have been studied extensively with images and videos, usually for deep learning and with generative adversarial networks. In the network case of interest to the Army, much less data is available and most classifiers in use are support-vector machines rather than neural networks. Hence, the focus is on support-vector machines. This project leverages ARL’s unique network emulation and data sets and focuses on techniques for constructing traffic that evades attack detection. Playing the adversary, the researchers were able to increase the false negative rate from 0 percent to 24 percent. Moving forward, the challenge will be to move to other kinds of classifiers (e.g., deep neural networks). This work was done in collaboration with university researchers.

Active Defense Systems for Vehicle Platforms Dynamic Watermarking

This research focuses on a novel method for detecting cyberattacks on vehicles and other cyber-physical systems. The approach is to inject a minor but known noise signal into control data in a way that allows detection of the output effects (the “watermark”). If control or sensor data is later hacked, the expectation is that the watermark will no longer be detectable. This project is closely related to a larger effort in active defense, and is being conducted in collaboration with University of Texas, Austin, and Texas A&M. The researchers have validated the ideas in the laboratory and are working toward validation in real vehicles to ensure that the injected signal does not adversely affect the operation of the vehicle. The work has been published in a leading journal and additional publication is being pursued in addition to filing of patent disclosures. The work is both novel and highly relevant to Army needs; it is unlikely that these kinds of attacks will be addressed by commercial automotive manufacturers. The validation of the approach with real vehicles lends additional credence to the results.

Fog Computing and the Tactical Distributed Ledger

These research projects represent two approaches to ensuring robust operation of a “network of things” such as cameras, directories, storage nodes, and so on, where the individual things fail, are replaced, or have degraded performance. In particular, the focus is on cases where if a particular resource that fails has generated data and done partial computation, a replacement would be able to continue from where its predecessor left off. Additionally, the research seeks approaches wherein new nodes would be able to join the network and to ensure that only authorized nodes be able to join.

The tactical distributed ledger project leverages a combination of existing blockchain technology and other techniques to solve many of these problems. A traditional public key/certificate approach is used for authorization. The tactical distributed ledger stores state in multiple locations, and “smart contract” technology allows resources to advertise their application programming interfaces and permits nodes that require services from resources to be able to advertise their needs.

The fog computing project is at an early stage of framing the overall concept. There are interesting questions that need to be addressed that pertain to efficient management of computing resources. For instance, is it better to expend network bandwidth to ensure that the state is captured reliably at all times (needed only if the resource dies), or to have the application start computation over and deal with partial data? Are certain tasks better attuned to one approach? Can the approach be flexible so that when certain resources become scarce, the approach adapts? The fog can even have connectivity to a cloud, so that select portions of the data can be more robustly stored, and reached if necessary, with greater latency, from the cloud.

Geometric and Topological Structures in Complex Networks

This project focuses on the structure of graphs and the role that geometry and topology can play in explaining key features in networks, such as predicting small group evolution or finding sparse covers. Real network data is used to determine the efficacy of techniques from geometry, such as curvature, or techniques from topology, such as homology, to identify features that are invariant under reduction. One motivating example is determining whether there are coverage holes due to GPS-denied network of devices. The project would benefit from a careful consideration of the network environments of the future Army. If this work is generally applicable to all networks, then researchers in academia and industry would be pursuing similar approaches; such connections to the literature were not made clear. A number of papers have appeared in conferences devoted to this area, and the work fits squarely within the basic research mission of ARL.

Narrative Generation for Human-Robot Interactions

This research is directed at developing approaches by which a robot can determine what happened in a block of time that is represented by sequence of images, and based on that understanding, generate a narrative for a human teammate. The problem is distinguished from related work available in the literature by constraints that are particularly important to the Army, such as low-quality, noncanonical imagery. This project integrates other lines of work at ARL such as the development of ontologies for image recognition. The approach uses crowdsourcing to generate general models of how humans handle narrative generation.

Uncertainty-Aware Artificial Intelligence and Machine Learning

The goal of this project is to develop methods to express both epistemic and aleatory uncertainty³ in ML and AI. Most AI approaches focus on characterizing aleatory uncertainty but do not perform well

___________________

³ Aleatory variability is the natural randomness in a process. For discrete variables, the randomness is parameterized by the probability of each possible value. For continuous variables, the randomness is parameterized by the probability density function. Epistemic uncertainty is the scientific uncertainty in the model of the process. It is due to limited data and knowledge. The epistemic uncertainty is characterized by alternative models.

with limited training data. The strategy used is to indicate the degree to which classification results depend on prior versus observed data. The approach, which includes replacing the soft-maximum probabilities usually produced by deep neural nets with parameters for probability distributions, appears to be novel and far-reaching. Initial work has been published in top-tier conferences and journals. The work is being pursued in collaboration with researchers at three universities.

Challenges and Opportunities

In most of the projects, it was clear how the scope of the research was informed by the needs and unique challenges of the Army. The researchers were able to express a clear understanding of the degree to which their work overlaps research performed in academia and other research laboratories.

In relation to the work on fog computing and the tactical distributed ledger, there are continuing challenges in the blockchain approach that could be explored in future research. For example, there are challenges with network partitioning and healing or merging two blockchains. Other challenges include storing the entire blockchain and replication of computation across each node. The researchers clearly understand these problems. It would be important to cleanly separate requirements of the fog application from the features provided by blockchains so that different substrates can be tried.

Several research projects fall into the general area of human-robot/machine interactions, with the two principal threads: scene narrative generation for humans by robots and robot learning from human demonstrations. This body of work is ambitious and has the potential to disrupt the way human-robot interactions are considered for future battlefields. The portfolio of research would benefit from closer interaction between the two principal threads. For instance, there is the potential to use inverse reinforcement learning methods to aid in making narrative generation more personal and contextual. The work would also benefit from investigation into the sensitivity of human narrative to mission scenarios.

Machine learning rightly pervades multiple research topics. The applications of this technology to various research interests may or may not yield favorable results, and healthy skepticism by researchers would be prudent.

OVERALL QUALITY OF THE WORK

Information Sciences

The research work in IS was assessed to be generally of high scientific quality, but not uniformly so. The research portfolio represents an appropriate balance of theoretical and experimental work, and many of the more mature programs—but not all—show a transition into practice or use by other areas.

In most cases, the research projects reflected a good understanding of the problems being considered, an appropriate statement of the problem being pursued, a good knowledge of the appropriate methodologies to address the problem, and acquaintance with the state of the art and the relevant research pursued elsewhere.

In many cases, the researchers are able to articulate Army relevance and identify research challenges that are unique to the Army’s operational needs. The researchers are well-qualified to carry out the research problems that they are pursuing and follow rigorous research methodologies and practices. Many of the projects have already resulted in publications in highly visible journals and selective conferences. The computation facilities and instrumentation required are adequate to the needs of the researchers.

Networks and Cyber

The overall scientific quality of the research was high, and comparable to research conducted at top research universities and government and industry laboratories. Researchers were very familiar with the underlying science and relevant leading research published and performed elsewhere. In many cases, there are active communications or collaborations with researchers outside ARL, many at the forefront of their respective fields. In all cases, the researchers were aware of the potential challenges, risks, and risk mitigations associated with their projects. In most cases, the researchers were able to incorporate these challenges, risks, and risk mitigations into their research.

There is an appropriate balance of theory, computational work, and physical experimentation to inform and investigate multiple areas of research. The facilities and supporting infrastructure are well-suited for collaborative work. There is a good mix of well-trained research personnel who also collaborate with researchers in a broad range of academic and industrial partners in addition to working with the ARL South and ARL West regional sites. The research staff in this area have received outstanding recognitions and awards for their research and technical contributions.

RECOMMENDATIONS

Information Sciences

The research related to imitation was found to be particularly noteworthy, drawing upon the notion that autonomous agents can learn via imitation of human “teachers,” of which the Army has many. The two approaches (reinforcement learning and inverse reinforcement learning) are not new, but the research here addresses cutting edge technology, and the results are potentially disruptive. It represents a new way to automate. The demonstration of the virtual reality was also viewed positively, with the recognition that notably high-quality experience could be disruptive for situational awareness related to Army operations. ARL could leverage the platform as a test-bed for ideas such as integration of satellite perspective, multiperson teaming, and human/agent teaming, and so on.

Most of the other projects, while not deemed as disruptive (a high bar), represent good research efforts and also provide ARL with the opportunity to develop human capital with expertise in ML and AI. Some projects would benefit from more ambitious goals, such as not only matching human levels of performance but perhaps surpassing it. Example topical areas where this may apply include autonomous navigation and visual saliency.

Much of the basic research in IS stems from the essential research area (ERA) projects or from one of the Collaborative Technology/Research Alliances (CTAs and CRAs). It is noted that many strong projects involve personnel embedded outside the ARL campus and that the initiative and coordination required to support such collaboration is to be lauded. The quality of the early-career ARL researchers in the IS area is high. The deployment of co-ops and interns, as well the Open Campus Initiative, has led to the recruitment of a strong cadre of emerging researchers.

Since this review was restricted to a subset of ongoing research projects, it was difficult to identify potential gaps in the portfolio. The coverage of AI appeared to be somewhat uneven, perhaps an outcome of the process of selection for presentation at the review. A stronger focus in areas like adversarial learning, integration of simulation in ML, and security related to ML would have provided a better understanding of the scope of ongoing work. As an example, simulation is emerging as a prominent element in training AI/ML systems, where physics-based simulation engines provide data to train autonomous vehicles and robots. Additionally, many technical challenges in modern ML involve problems that have been investigated in computational science. The Computational and Information Sciences Directorate (CISD) is well-poised to catalyze important new work that involves collaborations between computational and information sciences, an area that does not currently appear to be pursued.

Other important areas that were not represented relate to trust in AI and a focus on “explainable AI” that facilitates human understanding of AI actions as well as helps to engender trust.

Significant attention is given to AI and ML in research programs across different portfolios at ARL. At present, the scope of these research efforts is rather narrowly focused on mission-critical tasks. Fundamental research issues related to innovative ML techniques, AI implementation in resource-constrained environments, and trust and security of AI systems must still be addressed on a broader scale. It is also important to recognize that it is suboptimal to seek algorithmic advances in these areas without due consideration of hardware developments that are taking place in parallel. Given the potential of a disruptive impact of these technologies on Army operations, it is important to develop a comprehensive and integrative research plan in this emerging area. These technologies can have a transformational impact on key elements of Army operations.

Networks and Cyber

The research portfolio in networks and cyber comprises high-quality projects that are well aligned to Army needs. The research is directed at communications modalities, human-robot/machine interactions, and a study of adversarial systems as they relate to cybersecurity. It is not surprising that ML and AI are at the core of many of the ongoing projects. The pursuit of collaborations with internal and external groups performing related research has yielded rich dividends in the research results. Additionally, the approach of using demonstrations to test new concepts brings enhanced visibility to the research program and helps attract early-career researchers into the effort.

In the area of human-robot interactions, advances in narrative generation are significant and have the potential of not only reducing soldier workload but also reducing the bandwidth requirement for tactical networks. The work related to active defense was considered to be exceptionally strong and provides a reduction in the cyber vulnerability of Army vehicles and other cyber-enabled systems.

The approach of having multiple teams address the same problem using different approaches represents a useful research strategy. It provides a direct comparison between the approaches, and the best ideas from each approach can ultimately be leveraged and applied to the common problem. An example of this approach is within the fog computing and tactical distributed ledger projects. Another important research methodology, especially for security issues, is to study systems that have both defense and offense techniques. Once a defense technique is developed, the researchers can then see if an attacker (whether the defense technique is known or not) can be detected by a surveillance system. The “adversarial learning” project is a good example of researchers working on breaking existing norms, with the goal of leveraging what they learn to build better defenses.

Researchers in the networks and cyber areas do not routinely access the computational expertise within their respective larger organization. The quality and impact of research would be enhanced through increased collaboration with computational experts within ARL to leverage their unique and local expertise.

The science of complex computer networks is not well established. In particular, there are issues related to adoption of different protocols on shared communication systems. Army missions call for deployment of mobile wireless networks that must operate within a contended environment. In the absence of analytical solutions, emulation of such complex environments provides researchers with insights into the design and understanding of networks dynamics. Researchers in the networks and cyber areas do not routinely access the computational expertise within their respective larger organization. The quality and impact of their work would be enhanced through increased collaboration with computational experts within ARL to leverage their unique and local expertise.