Dario D. Salvucci
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Dario D. Salvucci has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- Ali Shokoufandeh
Drexel University
5 shared publications - Duncan P. Brumby
University College London, London, United Kingdom
4 shared publications - Daniel Schwartz
Drexel University
3 shared publications - Frank J. Lee
Drexel University, Philadelphia, PA, USA
3 shared publications - Genevieve Dion
Drexel University
3 shared publications - Niels Anne Taatgen
University of Groningen
3 shared publications
Latest Publications
- 15 July 2026
- Published by Drexel University Libraries
Extracting spatiotemporal concentrations of discriminative features from driving simulator time series data
A time series is a sequence of repeated numerical measurements taken over time. We employed novel methods for isolating and leveraging discriminative time series features (i.e., sub-series highly indicative of a class of time series samples) that were recorded during a simulated driving assessment, called the Virtual Driving Test (VDT), in order to develop a pre-screen for the on-road exam (ORE); identifying which license applicants that were about to take the ORE were most likely to fail the ORE (possibly dangerously so), based on their recorded VDT time series data. Our central hypothesis was that these discriminative time series features tended to be concentrated through time and space (i.e., spatiotemporally). We used Iterative Section Reduction (ISR), an ensemble-building algorithm that leveraged recursive data ablation to test for concentrations of discriminative time series features of various lengths recorded throughout the VDT planned route. ISR was used to build multivariate time series classification (MTSC) ensembles by systematically training and evaluating many section-specific VDT time series classifiers, dubbed modules, to act as voters in the ISR bagging ensembles in order to predict ORE outcome (pass/fail). Our most reliable MTSC models were ISR ensembles employing k-Medoids Dynamic Barycentric Averaging (k-Medoids DBA) modules using Dynamic Coordinate Locally Aligned Warping (DCLAW) as the time series distance (TSD) function for numerically comparing VDT time series samples to one another and predicting which samples were recorded by license applicants highly likely to fail the ORE (when taking it for the first time); achieving a composite evaluation risk ratio of 17.46 (14.59, 20.33). This was roughly four times as reliable as many commercially available breathalyzers are at correctly detecting if an individual's blood alcohol concentration (BAC) is over the legal limit. More than that, our ISR ensembles uncovered concentrations of discriminative time series features in the VDT data that suggested distracted driving combined with speeding through traffic light controlled intersections, turns, and curves in the roadway (particularly near a simulated playground where a child chased a ball into the roadway) could reliably distinguish license applicants that were highly likely to fail the ORE, particularly those that posed a potential threat to themselves or others through lack of sufficient practice before attempting the ORE, from all other applicants.
- 15 July 2026
- Published by Drexel University Libraries
Extracting spatiotemporal concentrations of discriminative features from driving simulator time series data
A time series is a sequence of repeated numerical measurements taken over time. We employed novel methods for isolating and leveraging discriminative time series features (i.e., sub-series highly indicative of a class of time series samples) that were recorded during a simulated driving assessment, called the Virtual Driving Test (VDT), in order to develop a pre-screen for the on-road exam (ORE); identifying which license applicants that were about to take the ORE were most likely to fail the ORE (possibly dangerously so), based on their recorded VDT time series data. Our central hypothesis was that these discriminative time series features tended to be concentrated through time and space (i.e., spatiotemporally). We used Iterative Section Reduction (ISR), an ensemble-building algorithm that leveraged recursive data ablation to test for concentrations of discriminative time series features of various lengths recorded throughout the VDT planned route. ISR was used to build multivariate time series classification (MTSC) ensembles by systematically training and evaluating many section-specific VDT time series classifiers, dubbed modules, to act as voters in the ISR bagging ensembles in order to predict ORE outcome (pass/fail). Our most reliable MTSC models were ISR ensembles employing k-Medoids Dynamic Barycentric Averaging (k-Medoids DBA) modules using Dynamic Coordinate Locally Aligned Warping (DCLAW) as the time series distance (TSD) function for numerically comparing VDT time series samples to one another and predicting which samples were recorded by license applicants highly likely to fail the ORE (when taking it for the first time); achieving a composite evaluation risk ratio of 17.46 (14.59, 20.33). This was roughly four times as reliable as many commercially available breathalyzers are at correctly detecting if an individual's blood alcohol concentration (BAC) is over the legal limit. More than that, our ISR ensembles uncovered concentrations of discriminative time series features in the VDT data that suggested distracted driving combined with speeding through traffic light controlled intersections, turns, and curves in the roadway (particularly near a simulated playground where a child chased a ball into the roadway) could reliably distinguish license applicants that were highly likely to fail the ORE, particularly those that posed a potential threat to themselves or others through lack of sufficient practice before attempting the ORE, from all other applicants.
- Open Access
- 29 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 10 (4) , 1-25
Resource-Efficient Gesture Recognition through Convexified Attention EICS006
Wearable e-textile interfaces require gesture recognition capabilities but face severe constraints in power consumption, computational capacity, and form factor that make traditional deep learning impractical. While lightweight architectures like MobileNet improve efficiency, they still demand thousands of parameters, limiting deployment on textile-integrated platforms. We introduce a convexified attention mechanism for wearable applications that dynamically weights features while preserving convexity through nonexpansive simplex projection and convex loss functions. Unlike conventional attention mechanisms using non-convex softmax operations, our approach employs Euclidean projection onto the probability simplex combined with multi-class hinge loss, ensuring global convergence guarantees. Implemented on a textile-based capacitive sensor with four connection points, our approach achieves 100.00% accuracy on tap gestures and 100.00% on swipe gestures—consistent across 10-fold cross-validation and held-out test evaluation—while requiring only 120–360 parameters, a 97% reduction compared to conventional approaches. With sub-millisecond inference times (290–296 μ s) and minimal storage requirements (< 7KB), our method enables gesture interfaces directly within e-textiles without external processing. Our evaluation, conducted in controlled laboratory conditions with a single-user dataset, demonstrates feasibility for basic gesture interactions. Real-world deployment would require validation across multiple users, environmental conditions, and more complex gesture vocabularies. These results demonstrate how convex optimization can enable efficient on-device machine learning for textile interfaces.
- 29 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 10 (4) , 1-25
Open AccessResource-Efficient Gesture Recognition through Convexified Attention EICS006
Wearable e-textile interfaces require gesture recognition capabilities but face severe constraints in power consumption, computational capacity, and form factor that make traditional deep learning impractical. While lightweight architectures like MobileNet improve efficiency, they still demand thousands of parameters, limiting deployment on textile-integrated platforms. We introduce a convexified attention mechanism for wearable applications that dynamically weights features while preserving convexity through nonexpansive simplex projection and convex loss functions. Unlike conventional attention mechanisms using non-convex softmax operations, our approach employs Euclidean projection onto the probability simplex combined with multi-class hinge loss, ensuring global convergence guarantees. Implemented on a textile-based capacitive sensor with four connection points, our approach achieves 100.00% accuracy on tap gestures and 100.00% on swipe gestures—consistent across 10-fold cross-validation and held-out test evaluation—while requiring only 120–360 parameters, a 97% reduction compared to conventional approaches. With sub-millisecond inference times (290–296 μ s) and minimal storage requirements (< 7KB), our method enables gesture interfaces directly within e-textiles without external processing. Our evaluation, conducted in controlled laboratory conditions with a single-user dataset, demonstrates feasibility for basic gesture interactions. Real-world deployment would require validation across multiple users, environmental conditions, and more complex gesture vocabularies. These results demonstrate how convex optimization can enable efficient on-device machine learning for textile interfaces.
- Open Access
Resource-Efficient Gesture Recognition through Convexified Attention
Wearable e-textile interfaces require gesture recognition capabilities but face severe constraints in power consumption, computational capacity, and form factor that make traditional deep learning impractical. While lightweight architectures like MobileNet improve efficiency, they still demand thousands of parameters, limiting deployment on textile-integrated platforms. We introduce a convexified attention mechanism for wearable applications that dynamically weights features while preserving convexity through nonexpansive simplex projection and convex loss functions. Unlike conventional attention mechanisms using non-convex softmax operations, our approach employs Euclidean projection onto the probability simplex combined with multi-class hinge loss, ensuring global convergence guarantees. Implemented on a textile-based capacitive sensor with four connection points, our approach achieves 100.00\% accuracy on tap gestures and 100.00\% on swipe gestures -- consistent across 10-fold cross-validation and held-out test evaluation -- while requiring only 120--360 parameters, a 97\% reduction compared to conventional approaches. With sub-millisecond inference times (290--296$μ$s) and minimal storage requirements ($<$7KB), our method enables gesture interfaces directly within e-textiles without external processing. Our evaluation, conducted in controlled laboratory conditions with a single-user dataset, demonstrates feasibility for basic gesture interactions. Real-world deployment would require validation across multiple users, environmental conditions, and more complex gesture vocabularies. These results demonstrate how convex optimization can enable efficient on-device machine learning for textile interfaces.
Open AccessResource-Efficient Gesture Recognition through Convexified Attention
Wearable e-textile interfaces require gesture recognition capabilities but face severe constraints in power consumption, computational capacity, and form factor that make traditional deep learning impractical. While lightweight architectures like MobileNet improve efficiency, they still demand thousands of parameters, limiting deployment on textile-integrated platforms. We introduce a convexified attention mechanism for wearable applications that dynamically weights features while preserving convexity through nonexpansive simplex projection and convex loss functions. Unlike conventional attention mechanisms using non-convex softmax operations, our approach employs Euclidean projection onto the probability simplex combined with multi-class hinge loss, ensuring global convergence guarantees. Implemented on a textile-based capacitive sensor with four connection points, our approach achieves 100.00\% accuracy on tap gestures and 100.00\% on swipe gestures -- consistent across 10-fold cross-validation and held-out test evaluation -- while requiring only 120--360 parameters, a 97\% reduction compared to conventional approaches. With sub-millisecond inference times (290--296$μ$s) and minimal storage requirements ($<$7KB), our method enables gesture interfaces directly within e-textiles without external processing. Our evaluation, conducted in controlled laboratory conditions with a single-user dataset, demonstrates feasibility for basic gesture interactions. Real-world deployment would require validation across multiple users, environmental conditions, and more complex gesture vocabularies. These results demonstrate how convex optimization can enable efficient on-device machine learning for textile interfaces.
- 1 September 2025
- Published by Springer Nature
- p. 46-59
A Minimal Neural Network for Reproducible Gesture Recognition on Knitted Capacitive Touch Sensors
- 1 September 2025
- Published by Springer Nature
- p. 46-59
A Minimal Neural Network for Reproducible Gesture Recognition on Knitted Capacitive Touch Sensors
- 1 July 2024
- Published by Drexel University Libraries
Pattern recognition techniques to infer driver intentions
- 1 July 2024
- Published by Drexel University Libraries
Pattern recognition techniques to infer driver intentions
- 16 August 2023
- Published by Drexel University Libraries
Representation Learning on Sequential and Temporal Data Manifolds
Representation learning has gained significant attention in the machine learning community due to its ability to automatically extract useful features from raw data without the need for hand-crafted feature engineering. Although representation learning is widely applied to image data, it has received less attention in the context of sequential and temporal data manifolds, despite the growing prevalence of such data in industry and on the internet. This dissertation focuses on utilizing representation learning in applications with sequential and temporal data and demonstrates its advantages over conventional techniques that do not incorporate a sequential component. We investigate the different tools and techniques used in representation learning for sequential and temporal data, compare their strengths and weaknesses, and provide guidelines for implementing them in various domains. The study also presents a number of case studies in a variety of domains to demonstrate the benefits of representation learning. In the time series case study, representation learning is applied to analyze the driving behavior of individuals with cognitive impairments, demonstrating its effectiveness over traditional techniques. In the text and graph case study, representation learning is extended to clinical graphs to present disease information in an interpretable space, measuring patient similarity based on their visit notes. In the image case study, transfer learning is used to improve object classification models via neuro-inspired approaches to enhance representation learning for static image classification. This case study explores the creation of a dynamic perception framework that is robust to minor perturbations in input stimuli through representation learning over time. Among the contributions of this dissertation are a comprehensive review of representation learning techniques for sequential and temporal data manifolds. Also included are guidelines for their use in various domains with such data, as well as real-world case studies demonstrating their effectiveness. The purpose of this dissertation is to provide insight into how representation learning can improve models' performance, reduce data dimensionality, and facilitate models' interpretation.
- 16 August 2023
- Published by Drexel University Libraries
Representation Learning on Sequential and Temporal Data Manifolds
Representation learning has gained significant attention in the machine learning community due to its ability to automatically extract useful features from raw data without the need for hand-crafted feature engineering. Although representation learning is widely applied to image data, it has received less attention in the context of sequential and temporal data manifolds, despite the growing prevalence of such data in industry and on the internet. This dissertation focuses on utilizing representation learning in applications with sequential and temporal data and demonstrates its advantages over conventional techniques that do not incorporate a sequential component. We investigate the different tools and techniques used in representation learning for sequential and temporal data, compare their strengths and weaknesses, and provide guidelines for implementing them in various domains. The study also presents a number of case studies in a variety of domains to demonstrate the benefits of representation learning. In the time series case study, representation learning is applied to analyze the driving behavior of individuals with cognitive impairments, demonstrating its effectiveness over traditional techniques. In the text and graph case study, representation learning is extended to clinical graphs to present disease information in an interpretable space, measuring patient similarity based on their visit notes. In the image case study, transfer learning is used to improve object classification models via neuro-inspired approaches to enhance representation learning for static image classification. This case study explores the creation of a dynamic perception framework that is robust to minor perturbations in input stimuli through representation learning over time. Among the contributions of this dissertation are a comprehensive review of representation learning techniques for sequential and temporal data manifolds. Also included are guidelines for their use in various domains with such data, as well as real-world case studies demonstrating their effectiveness. The purpose of this dissertation is to provide insight into how representation learning can improve models' performance, reduce data dimensionality, and facilitate models' interpretation.
- 16 July 2021
- Published by Drexel University Libraries
Geometric and Combinatorial Methods for Digital Elevation Models
Digital Elevation Models (DEMs) are the de facto 3D models used to represent geo-spatial structures in many natural systems and are used in process-based modeling. From a computational perspective, DEMs are discrete, 2D manifold-grids in 3D space from which flow properties can be computed. An important family of problems involves identifying iso-flow surfaces from DEMs at discrete locations. In practical terms, such iso-flow surfaces correspond to sub-regions of the 3D-manifold whose aggregate flow passesthrough a single grid cell. Algorithmic solutions to these types of problems have numerous applications in Geographic Information Systems and Environmental Science research, allowing the efficient mapping of watershed boundaries. Existing algorithms for identifying iso-surfaces run in quadratic complexity as a function of the number of cells on their boundary. Our investigation shows that linear time algorithms to identify iso-surfaces are possible due to the combinatorial geometric properties of DEMs. Finally, a general method to return univariate statistical values for an iso-flow surfaces is developed. These algorithms rely on a number of new data structures, the Modified Nested Sets, Log Reduced Nested Sets, and a Modified Interval Search Tree. In empirical tests we show significant ( 500 faster) for a number of these iso-surface problems.
- 16 July 2021
- Published by Drexel University Libraries
Geometric and Combinatorial Methods for Digital Elevation Models
Digital Elevation Models (DEMs) are the de facto 3D models used to represent geo-spatial structures in many natural systems and are used in process-based modeling. From a computational perspective, DEMs are discrete, 2D manifold-grids in 3D space from which flow properties can be computed. An important family of problems involves identifying iso-flow surfaces from DEMs at discrete locations. In practical terms, such iso-flow surfaces correspond to sub-regions of the 3D-manifold whose aggregate flow passesthrough a single grid cell. Algorithmic solutions to these types of problems have numerous applications in Geographic Information Systems and Environmental Science research, allowing the efficient mapping of watershed boundaries. Existing algorithms for identifying iso-surfaces run in quadratic complexity as a function of the number of cells on their boundary. Our investigation shows that linear time algorithms to identify iso-surfaces are possible due to the combinatorial geometric properties of DEMs. Finally, a general method to return univariate statistical values for an iso-flow surfaces is developed. These algorithms rely on a number of new data structures, the Modified Nested Sets, Log Reduced Nested Sets, and a Modified Interval Search Tree. In empirical tests we show significant ( 500 faster) for a number of these iso-surface problems.
- 16 July 2021
- Published by Drexel University Libraries
Canonical behavior patterns
A common problem in many areas of behavioral research is the analysis of the large volume of protocol data recorded during the execution of tasks. This dissertation describes a new automated method of protocol analysis to find canonical behaviors: a small subset of behavior protocols that are most representative of the full data set. The method I have developed takes advantage of recent algorithmic developments in pattern recognition. By adapting these methods to the analysis of behavior protocols, I provide a new tool for analysts working with large datasets that are infeasible to study using current methods. The method I propose can also be used as an important complement to existing sequential protocol analysis techniques, by allowing researchers to build their models based on a few highly representative samples. The contributions of this dissertation include the adaptation of the method to the analysis of behavior protocols; the development of similarity measures appropriate to behavior protocols; an extension of the method to work in oriented topologies; and a demonstration of the method's utility in real-world problem domains, particularly web browsing and driving.
- 16 July 2021
- Published by Drexel University Libraries
Canonical behavior patterns
A common problem in many areas of behavioral research is the analysis of the large volume of protocol data recorded during the execution of tasks. This dissertation describes a new automated method of protocol analysis to find canonical behaviors: a small subset of behavior protocols that are most representative of the full data set. The method I have developed takes advantage of recent algorithmic developments in pattern recognition. By adapting these methods to the analysis of behavior protocols, I provide a new tool for analysts working with large datasets that are infeasible to study using current methods. The method I propose can also be used as an important complement to existing sequential protocol analysis techniques, by allowing researchers to build their models based on a few highly representative samples. The contributions of this dissertation include the adaptation of the method to the analysis of behavior protocols; the development of similarity measures appropriate to behavior protocols; an extension of the method to work in oriented topologies; and a demonstration of the method's utility in real-world problem domains, particularly web browsing and driving.
- 16 July 2021
- Published by Drexel University Libraries
Mobile intelligent tutoring system
The prominence of computers in the 21st Century has caused educators to reexamine the needs of today's K-12 students. The proposed 21st Century Skills center on the use of technology and thus millions of students are unable fully to master these skills due to their minimal access to technology. The Digital Divide, a correlation between an individual's access to technology and their socio-economic status, is poised to prevent large segments of society from advancing and thriving in this technology based economy. One proposed solution to this lack of computing resources lay in the transformation of mobile devices, like cell phones, from single-purpose communication tools to multipurpose computing resources. The smaller scale of mobile devices mandates the design of mobile learning applications be more than miniature desktop learning applications. Whether the applications are based on existing desktop applications or presenting new paradigms of learning activities there are human-computer-interaction and education concerns to address. This thesis describes the design, implementation, and evaluation of a complete mobile intelligent tutoring system (ITS). The research questions investigated address the design and evaluation of the implemented system. To answer first question, "How might the design of an ITS be adapted for delivery on a mobile device?" a set of general mobile ITS guidelines is described. To answer the second question, "Can a mobile ITS provide learning gains greater than traditional instructional activities?" data was gathered from a controlled study to compare the performance of students using the mobile ITS with those who did not. These data reveal that students using the mobile ITS achieved gains greater than students receiving standard instruction. The third question, "What teaching strategy best supports a mobile ITS?", is answered through data gathered from the comparison of two teaching strategies, long and short. These data suggest that the long strategy, in which students are required to calculate the final answer to the problem, best facilitates student performance gains. This thesis lays the foundation for future research that explores the delivery of intelligent tutoring systems off the desktop as well as for research in methods of transforming desktop learning applications for mobile devices.
- 16 July 2021
- Published by Drexel University Libraries
Mobile intelligent tutoring system
The prominence of computers in the 21st Century has caused educators to reexamine the needs of today's K-12 students. The proposed 21st Century Skills center on the use of technology and thus millions of students are unable fully to master these skills due to their minimal access to technology. The Digital Divide, a correlation between an individual's access to technology and their socio-economic status, is poised to prevent large segments of society from advancing and thriving in this technology based economy. One proposed solution to this lack of computing resources lay in the transformation of mobile devices, like cell phones, from single-purpose communication tools to multipurpose computing resources. The smaller scale of mobile devices mandates the design of mobile learning applications be more than miniature desktop learning applications. Whether the applications are based on existing desktop applications or presenting new paradigms of learning activities there are human-computer-interaction and education concerns to address. This thesis describes the design, implementation, and evaluation of a complete mobile intelligent tutoring system (ITS). The research questions investigated address the design and evaluation of the implemented system. To answer first question, "How might the design of an ITS be adapted for delivery on a mobile device?" a set of general mobile ITS guidelines is described. To answer the second question, "Can a mobile ITS provide learning gains greater than traditional instructional activities?" data was gathered from a controlled study to compare the performance of students using the mobile ITS with those who did not. These data reveal that students using the mobile ITS achieved gains greater than students receiving standard instruction. The third question, "What teaching strategy best supports a mobile ITS?", is answered through data gathered from the comparison of two teaching strategies, long and short. These data suggest that the long strategy, in which students are required to calculate the final answer to the problem, best facilitates student performance gains. This thesis lays the foundation for future research that explores the delivery of intelligent tutoring systems off the desktop as well as for research in methods of transforming desktop learning applications for mobile devices.
- 16 July 2021
- Published by Drexel University Libraries
A Unified Model of Fatigue in a Cognitive Architecture
Capturing the effects of fatigue and, more generally, the effects of physical and mental states on human performance has been a topic of research for many years. Recent models, especially those developed in a cognitive architecture, have shown great promise in capturing these effects by providing insight into the specific cognitive and other components involved in task performance (like perception and motor movement). In particular, separate models have been developed to account for both time-of-day (sleep deprivation, sleep restriction, and circadian rhythms) and time-on-task effects related to fatigue. Unfortunately, a unified model to capture these aspects of fatigue has been elusive. This dissertation presents a novel unified model, developed in the ACT-R cognitive architecture, that captures both time-of-day and time-on-task effects with a single set of mechanisms and parameters. The thesis demonstrates how this unified model accounts for quantitative and qualitative aspects of fatigued performance from a number of experimental studies. First, the thesis shows how the model captures behavior in the common psychomotor vigilance task (PVT), examining time-on-task effects under conditions of moderate fatigue as well as time-of-day effects under conditions of severe long-term (88-hour) sleep deprivation. Second, the thesis demonstrates how the model accounts for behavior when combining PVT with the more complex task of driving, using recent studies: one that focuses on driving under conditions of moderate fatigue, caused by repeated periods of nighttime driving, in a simulated environment; and another four-hour driving experiment in a real driving scenario focused on time-on-task effects on driving. Overall, the thesis demonstrates that the unified model can, with the same mechanisms and set of parameters, account for a range of effects of fatigue that shows promise to generalizing to other cognitive tasks.
- 16 July 2021
- Published by Drexel University Libraries
A Unified Model of Fatigue in a Cognitive Architecture
Capturing the effects of fatigue and, more generally, the effects of physical and mental states on human performance has been a topic of research for many years. Recent models, especially those developed in a cognitive architecture, have shown great promise in capturing these effects by providing insight into the specific cognitive and other components involved in task performance (like perception and motor movement). In particular, separate models have been developed to account for both time-of-day (sleep deprivation, sleep restriction, and circadian rhythms) and time-on-task effects related to fatigue. Unfortunately, a unified model to capture these aspects of fatigue has been elusive. This dissertation presents a novel unified model, developed in the ACT-R cognitive architecture, that captures both time-of-day and time-on-task effects with a single set of mechanisms and parameters. The thesis demonstrates how this unified model accounts for quantitative and qualitative aspects of fatigued performance from a number of experimental studies. First, the thesis shows how the model captures behavior in the common psychomotor vigilance task (PVT), examining time-on-task effects under conditions of moderate fatigue as well as time-of-day effects under conditions of severe long-term (88-hour) sleep deprivation. Second, the thesis demonstrates how the model accounts for behavior when combining PVT with the more complex task of driving, using recent studies: one that focuses on driving under conditions of moderate fatigue, caused by repeated periods of nighttime driving, in a simulated environment; and another four-hour driving experiment in a real driving scenario focused on time-on-task effects on driving. Overall, the thesis demonstrates that the unified model can, with the same mechanisms and set of parameters, account for a range of effects of fatigue that shows promise to generalizing to other cognitive tasks.