Symbolic Deep Networks: A Psychologically Inspired Lightweight and Efficient Approach to Deep Learning.

The last two decades have produced unprecedented successes in the fields of artificial intelligence and machine learning (ML), due almost entirely to advances in deep neural networks (DNNs). Deep hierarchical memory networks are not a novel concept in cognitive science and can be traced back more than a half century to Simon's early work on discrimination nets for simulating human expertise. The major difference between DNNs and the deep memory nets meant for explaining human cognition is that the latter are symbolic networks meant to model the dynamics of human memory and learning. Cognition-inspired symbolic deep networks (SDNs) address several known issues with DNNs, including (1) learning efficiency, where a much larger number of training examples are required for DNNs than would be expected for a human; (2) catastrophic interference, where what is learned by a DNN gets unlearned when a new problem is presented; and (3) explainability, where there is no way to explain what is learned by a DNN. This paper explores whether SDNs can achieve similar classification accuracy performance to DNNs across several popular ML datasets and discusses the strengths and weaknesses of each approach. Simulations reveal that (1) SDNs provide similar accuracy to DNNs in most cases, (2) SDNs are far more efficient than DNNs, (3) SDNs are as robust as DNNs to irrelevant/noisy attributes in the data, and (4) SDNs are far more robust to catastrophic interference than DNNs. We conclude that SDNs offer a promising path toward human-level accuracy and efficiency in category learning. More generally, ML frameworks could stand to benefit from cognitively inspired approaches, borrowing more features and functionality from models meant to simulate and explain human learning.

Cognitive Models in Cybersecurity: Learning From Expert Analysts and Predicting Attacker Behavior.

Cybersecurity stands to benefit greatly from models able to generate predictions of attacker and defender behavior. On the defender side, there is promising research suggesting that Symbolic Deep Learning (SDL) may be employed to automatically construct cognitive models of expert behavior based on small samples of expert decisions. Such models could then be employed to provide decision support for non-expert users in the form of explainable expert-based suggestions. On the attacker side, there is promising research suggesting that model-tracing with dynamic parameter fitting may be used to automatically construct models during live attack scenarios, and to predict individual attacker preferences. Predicted attacker preferences could then be exploited for mitigating risk of successful attacks. In this paper we examine how these two cognitive modeling approaches may be useful for cybersecurity professionals via two human experiments. In the first experiment participants play the role of cyber analysts performing a task based on Intrusion Detection System alert elevation. Experiment results and analysis reveal that SDL can help to reduce missed threats by 25%. In the second experiment participants play the role of attackers picking among four attack strategies. Experiment results and analysis reveal that model-tracing with dynamic parameter fitting can be used to predict (and exploit) most attackers' preferences 40-70% of the time. We conclude that studies and models of human cognition are highly valuable for advancing cybersecurity.

Participation shifts explain degree distributions in a human communications network.

Human interpersonal communications drive political, technological, and economic systems, placing importance on network link prediction as a fundamental problem of the sciences. These systems are often described at the network-level by degree counts -the number of communication links associated with individuals in the network-that often follow approximate Pareto distributions, a divergence from Poisson-distributed counts associated with random chance. A defining challenge is to understand the inter-personal dynamics that give rise to such heavy-tailed degree distributions at the network-level; primarily, these distributions are explained by preferential attachment, which, under certain conditions, can create power law distributions; preferential attachment's prediction of these distributions breaks down, however, in conditions with no network growth. Analysis of an organization's email network suggests that these degree distributions may be caused by the existence of individual participation-shift dynamics that are necessary for coherent communication between humans. We find that the email network's degree distribution is best explained by turn-taking and turn-continuing norms present in most social network communication. We thus describe a mechanism to explain a long-tailed degree distribution in conditions with no network growth.

Cyber Teaming and Role Specialization in a Cyber Security Defense Competition.

A critical requirement for developing a cyber capable workforce is to understand how to challenge, assess, and rapidly develop human cyber skill-sets in realistic cyber operational environments. Fortunately, cyber team competitions make use of simulated operational environments with scoring criteria of task performance that objectively define overall team effectiveness, thus providing the means and context for observation and analysis of cyber teaming. Such competitions allow researchers to address the key determinants that make a cyber defense team more or less effective in responding to and mitigating cyber attacks. For this purpose, we analyzed data collected at the 12th annual Mid-Atlantic Collegiate Cyber Defense Competition (MACCDC, http://www.maccdc.org), where eight teams were evaluated along four independent scoring dimensions: maintaining services, incident response, scenario injects, and thwarting adversarial activities. Data collected from the 13-point OAT (Observational Assessment of Teamwork) instrument by embedded observers and a cyber teamwork survey completed by all participants were used to assess teamwork and leadership behaviors and team composition and work processes, respectively. The scores from the competition were used as an outcome measure in our analysis to extract key features of team process, structure, leadership, and skill-sets in relation to effective cyber defense. We used Bayesian regression to relate scored performance during the competition to team skill composition, team experience level, and an observational construct of team collaboration. Our results indicate that effective collaboration, experience, and functional role-specialization within the teams are important factors that determine the success of these teams in the competition and are important observational predictors of the timely detection and effective mitigation of ongoing cyber attacks. These results support theories of team maturation and the development of functional team cognition applied to mastering cybersecurity.

Simulations in Cyber-Security: A Review of Cognitive Modeling of Network Attackers, Defenders, and Users.

Computational models of cognitive processes may be employed in cyber-security tools, experiments, and simulations to address human agency and effective decision-making in keeping computational networks secure. Cognitive modeling can addresses multi-disciplinary cyber-security challenges requiring cross-cutting approaches over the human and computational sciences such as the following: (a) adversarial reasoning and behavioral game theory to predict attacker subjective utilities and decision likelihood distributions, (b) human factors of cyber tools to address human system integration challenges, estimation of defender cognitive states, and opportunities for automation, (c) dynamic simulations involving attacker, defender, and user models to enhance studies of cyber epidemiology and cyber hygiene, and (d) training effectiveness research and training scenarios to address human cyber-security performance, maturation of cyber-security skill sets, and effective decision-making. Models may be initially constructed at the group-level based on mean tendencies of each subject's subgroup, based on known statistics such as specific skill proficiencies, demographic characteristics, and cultural factors. For more precise and accurate predictions, cognitive models may be fine-tuned to each individual attacker, defender, or user profile, and updated over time (based on recorded behavior) via techniques such as model tracing and dynamic parameter fitting.

Enhancing Team Composition in Professional Networks: Problem Definitions and Fast Solutions.

In this paper, we study ways to enhance the composition of teams based on new requirements in a collaborative environment. We focus on recommending team members who can maintain the team's performance by minimizing changes to the team's skills and social structure. Our recommendations are based on computing team-level similarity, which includes skill similarity, structural similarity as well as the synergy between the two. Current heuristic approaches are one-dimensional and not comprehensive, as they consider the two aspects independently. To formalize team-level similarity, we adopt the notion of graph kernel of attributed graphs to encompass the two aspects and their interaction. To tackle the computational challenges, we propose a family of fast algorithms by (a) designing effective pruning strategies, and (b) exploring the smoothness between the existing and the new team structures. Extensive empirical evaluations on real world datasets validate the effectiveness and efficiency of our algorithms.

Mission Command in the Age of Network-Enabled Operations: Social Network Analysis of Information Sharing and Situation Awareness.

A common assumption in organizations is that information sharing improves situation awareness and ultimately organizational effectiveness. The sheer volume and rapid pace of information and communications received and readily accessible through computer networks, however, can overwhelm individuals, resulting in data overload from a combination of diverse data sources, multiple data formats, and large data volumes. The current conceptual framework of network enabled operations (NEO) posits that robust networking and information sharing act as a positive feedback loop resulting in greater situation awareness and mission effectiveness in military operations (Alberts and Garstka, 2004). We test this assumption in a large-scale, 2-week military training exercise. We conducted a social network analysis of email communications among the multi-echelon Mission Command staff (one Division and two sub-ordinate Brigades) and assessed the situational awareness of every individual. Results from our exponential random graph models challenge the aforementioned assumption, as increased email output was associated with lower individual situation awareness. It emerged that higher situation awareness was associated with a lower probability of out-ties, so that broadly sending many messages decreased the likelihood of attaining situation awareness. This challenges the hypothesis that increased information sharing improves situation awareness, at least for those doing the bulk of the sharing. In addition, we observed two trends that reflect a compartmentalizing of networked information sharing as email links were more commonly formed among members of the command staff with both similar functions and levels of situation awareness, than between two individuals with dissimilar functions and levels of situation awareness; both those findings can be interpreted to reflect effects of homophily. Our results have major implications that challenge the current conceptual framework of NEO. In addition, the information sharing network was largely imbalanced and dominated by a few key individuals so that most individuals in the network have very few email connections, but a small number of individuals have very many connections. These results highlight several major growing pains for networked organizations and military organizations in particular.

Effects of Information Availability on Command-and-Control Decision Making: Performance, Trust, and Situation Awareness.

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.

Where is ELSA? The early to late shift in aging.

Studies of cognitive and neural aging have recently provided evidence of a shift from an early- to late-onset cognitive control strategy, linked with temporally extended activity in the prefrontal cortex (PFC). It has been uncertain, however, whether this age-related shift is unique to PFC and executive control tasks or whether the functional location might vary depending on the particular cognitive processes that are altered. The present study tested whether an early-to-late shift in aging (ELSA) might emerge in the medial temporal lobes (MTL) during a protracted context memory task comprising both anticipatory cue (retrieval preparation) and retrieval probe (retrieval completion) phases. First, we found reduced MTL activity in older adults during the early retrieval preparation phase coupled with increased MTL activity during the late retrieval completion phase. Second, we found that functional connectivity between MTL and PFC regions was higher during retrieval preparation in young adults but higher during retrieval completion in older adults, suggesting an important interactive relationship between the ELSA pattern in MTL and PFC. Taken together, these results critically suggest that aging results in temporally lagged activity even in regions not typically associated with cognitive control, such as the MTL.

Neural correlates of confidence during item recognition and source memory retrieval: evidence for both dual-process and strength memory theories.

Although the medial-temporal lobes (MTL), PFC, and parietal cortex are considered primary nodes in the episodic memory network, there is much debate regarding the contributions of MTL, PFC, and parietal subregions to recollection versus familiarity (dual-process theory) and the feasibility of accounts on the basis of a single memory strength process (strength theory). To investigate these issues, the current fMRI study measured activity during retrieval of memories that differed quantitatively in terms of strength (high vs. low-confidence trials) and qualitatively in terms of recollection versus familiarity (source vs. item memory tasks). Support for each theory varied depending on which node of the episodic memory network was considered. Results from MTL best fit a dual-process account, as a dissociation was found between a right hippocampal region showing high-confidence activity during the source memory task and bilateral rhinal regions showing high-confidence activity during the item memory task. Within PFC, several left-lateralized regions showed greater activity for source than item memory, consistent with recollective orienting, whereas a right-lateralized ventrolateral area showed low-confidence activity in both tasks, consistent with monitoring processes. Parietal findings were generally consistent with strength theory, with dorsal areas showing low-confidence activity and ventral areas showing high-confidence activity in both tasks. This dissociation fits with an attentional account of parietal functions during episodic retrieval. The results suggest that both dual-process and strength theories are partly correct, highlighting the need for an integrated model that links to more general cognitive theories to account for observed neural activity during episodic memory retrieval.

A search-by-clusters model of visual search: fits to data from younger and older adults.

OBJECTIVES: This study aims to specify the processing operations underlying age-related differences in the speed and accuracy of visual search in a mathematical model. METHOD: Eighteen older and 18 young adults searched for a predesignated target within 24-degree visual arrays containing distractors. Targets were systematically placed in regions that extended 2.5, 5.0, 7.5, and 10 degrees from center. Data were fitted to several versions of a mathematical model in which it was assumed that target search proceeds from the center fixation to peripheral areas in a succession of visual inspections of clusters until the target is located and that clusters can vary in size in response to search difficulty. RESULTS: Eccentricity effects on latencies and errors were larger for older adults than for younger adults, especially in the hardest search condition. The best-fitting version of the "search-by-clusters" model accounted for an average of 98.4% and 95.4% of the variance in the young and older adults, respectively. The resulting time, accuracy, and cluster parameters behaved plausibly in each of the 36 data sets. CONCLUSIONS: A quantitative model that specified how individuals searched for targets in large arrays accurately predicted the search times and accuracies of younger and older adults.

Effects of repetition on associative recognition in young and older adults: item and associative strengthening.

Young and older adults studied word pairs and later discriminated studied pairs from various types of foils including recombined word-pairs and foil pairs containing one or two previously unstudied words. We manipulated how many times a specific word pair was repeated (1 or 5) and how many different words were associated with a given word (1 or 5) to tease apart the effects of item familiarity from recollection of the association. Rather than making simple old/new judgments, subjects chose one of five responses: (a) Old-Old (original), (b) Old-Old (rearranged), (c) Old-New, (d) New-Old, (e) New-New. Veridical recollection was impaired in old age in all memory conditions. There was evidence for a higher rate of false recollection of rearranged pairs following exact repetition of study pairs in older but not younger adults. In contrast, older adults were not more susceptible to interference than young adults when one or both words of the pair had multiple competing associates. Older adults were just as able as young adults to use item familiarity to recognize which word of a foil was old. This pattern suggests that recollection problems in advanced age are because of a deficit in older adults' formation or retrieval of new associations in memory. A modeling simulation provided good fits to these data and offers a mechanistic explanation based on an age-related reduction of working memory.

Assessing the effects of age on long white matter tracts using diffusion tensor tractography.

Aging is associated with significant white matter deterioration and this deterioration is assumed to be at least partly a consequence of myelin degeneration. The present study investigated specific predictions of the myelodegeneration hypothesis using diffusion tensor tractography. This technique has several advantages over other methods of assessing white matter architecture, including the possibility of isolating individual white matter tracts and measuring effects along the whole extent of each tract. The study yielded three main findings. First, age-related white matter deficits increased gradually from posterior to anterior segments within specific fiber tracts traversing frontal and parietal, but not temporal cortex. This pattern inverts the sequence of myelination during childhood and early development observed in previous studies and lends support to a "last-in-first-out" theory of the white matter health across the lifespan. Second, both the effects of aging on white matter and their impact on cognitive performance were stronger for radial diffusivity (RD) than for axial diffusivity (AD). Given that RD has previously been shown to be more sensitive to myelin integrity than AD, this second finding is also consistent with the myelodegeneration hypothesis. Finally, the effects of aging on select white matter tracts were associated with age difference in specific cognitive functions. Specifically, FA in anterior tracts was shown to be primarily associated with executive tasks and FA in posterior tracts mainly associated with visual memory tasks. Furthermore, these correlations were mirrored in RD, but not AD, suggesting that RD is more sensitive to age-related changes in cognition. Taken together, the results help to clarify how age-related white matter decline impairs cognitive performance.

Memory for Items and Associations: Distinct Representations and Processes in Associative Recognition.

In two experiments, participants studied word pairs and later discriminated old (intact) word pairs from foils, including recombined word pairs and pairs including one or two previously unstudied words. Rather than making old/new memory judgments, they chose one of five responses: (1) Old-Old (original), (2) Old-Old (rearranged), (3) Old-New, (4) New-Old, (5) New-New. To tease apart the effects of item familiarity from those of associative strength, we varied both how many times a specific word-pair was repeated (1 or 5) and how many different word pairs were associated with a given word (1 or 5). Participants could discriminate associative information from item information such that they recognized which word of a foil was new, or whether both were new, as well as discriminating recombined studied words from original pairings. The error and latency data support the view that item and associative information are stored as distinct memory representations and make separate contributions at retrieval.

Rules and more rules: the effects of multiple tasks, extensive training, and aging on task-switching performance.

Task-switching performance was assessed in young and older adults as a function of the number of task sets to be actively maintained in memory (varied from 1 to 4) over the course of extended training (5 days). Each of the four tasks required the execution of a simple computational algorithm, which was instantaneously cued by the color of the two-digit stimulus. Tasks were presented in pure (task set size 1) and mixed blocks (task set sizes 2, 3, 4), and the task sequence was unpredictable. By considering task switching beyond two tasks, we found evidence for a cognitive control system that is not overwhelmed by task set size load manipulations. Extended training eliminated age effects in task-switching performance, even when the participants had to manage the execution of up to four tasks. The results are discussed in terms of current theories of cognitive control, including task set inertia and production system postulates.

Acute alcohol intoxication impairs controlled search across the visual field.

OBJECTIVE: The objective of the two experiments reported was to examine the effects of acute alcohol intoxication and of alcohol expectancies on controlled search across the visual field. METHOD: After receiving an oral dose of 0.5 g/kg alcohol (Experiment 1), an oral dose of 0.7 g/kg alcohol (Experiment 2), or a placebo, participants searched for a target in large arrays of homogeneous distractors that were either highly similar or less similar to the target. Targets were systematically placed at fixation and at visual angles of 2.5 degrees , 5.0 degrees , 7.5 degrees , and 10.0 degrees . RESULTS: Target detection was less accurate in the placebo condition of both experiments than in the no-beverage control group, suggesting that alcohol expectancies had a negative effect on controlled search. The effects of alcohol at the lower dose and of the placebo on visual search were not different (Experiment 1). At the higher dose, the negative effects of target eccentricity on the accuracy of target detection were larger when targets appeared among highly similar distractors, compared with the placebo condition and with a no-beverage control group. Target eccentricity effects on accuracy or speed were not observed at either dose when targets and distractors were dissimilar. CONCLUSIONS: Acute alcohol intoxication at either a low or high dose and placebo-associated intoxication expectancies have a detrimental effect on controlled visual search in large arrays. Acute alcohol intoxication at a high dosage exaggerates the detrimental effects of target eccentricity and of task difficulty on controlled visual search.

Modeling age-related memory deficits: a two-parameter solution.

Recent research and meta-analytic reviews suggest that 1 observed pattern of impaired and intact memory performance with advancing age is a deficit in measures of episodic but not semantic memory. The authors used computational modeling to explore a number of age-related parameters to account for this pattern. A 2-parameter solution based on lifelong experience successfully fit the pattern of results in 5 published studies of the word-frequency mirror effect and paired-associate recognition. Lifelong experience increases the strength (resting level of activation) of concepts in the network but also saturates the network with an increasing number of episodic associations to each concept. More episodic associations to each concept mean that activation spreads more diffusely, making retrieval of any newly established memory trace less likely; however, the greater strength of a concept makes recognition based on familiarity more likely. The simulations provide good quantitative fits to the extant age-related memory literature and support the plausibility of this mechanistic account.