Research News

In the ever-evolving landscape of urban transportation, ride-pooling services have emerged as a promising solution, offering a shared mobility experience that is both cost-effective and environmentally friendly. However, optimizing these services to achieve high ridership while maintaining efficiency has remained a challenge. A new study by NYU Trandon transportation experts proposes a novel algorithm aimed at revolutionizing the way ride-pooling services operate.

Led by Joseph Chow, Institute Associate Professor of Civil and Urban Engineering and Deputy Director of the USDOT Tier 1 University Transportation Center C2SMARTER, the study delves into the intricacies of dynamic routing in ride-pooling services, with a particular focus on the integration of transfers within the system. Transfers, the process of passengers switching between vehicles during their journey, have long been identified as a potential strategy to enhance service availability and fleet efficiency. Yet, the implementation of transfers poses a highly complex routing problem, one that has largely been overlooked in existing literature.

The research team's solution comes in the form of a state-of-the-art dynamic routing algorithm, designed to incorporate synchronized intramodal transfers seamlessly into the ride-pooling experience. Unlike traditional approaches that focus solely on immediate decisions, the proposed algorithm adopts a forward-looking perspective, taking into account the long-term implications of routing choices.

Central to the study is the development of a simulation platform, allowing researchers to implement and test their proposed algorithm in real-world scenarios. Drawing on data from both the Sioux Falls network and the MOIA ride-pooling service in Hamburg, Germany, the team evaluated the performance of their algorithm across various operational settings.

The results of the study are promising, suggesting that the incorporation of transfers into ride-pooling services using the proposed algorithm can lead to significant improvements in fleet utilization and service quality compared to transfers without it. By accounting for the additional opportunity costs of transfer commitments, the proposed algorithm demonstrates a competitive edge over traditional myopic approaches, reducing operating costs per passenger and minimizing the number of rejected ride requests.

While the findings represent a significant advancement in the field of urban transportation, the researchers acknowledge that further validation and refinement are necessary before widespread implementation. Nonetheless, the study marks a pivotal moment in the ongoing quest to optimize ride-pooling services for the cities of tomorrow.

In summary, the contributions of the research can be categorized into three main areas:

1. Development of an innovative online policy and algorithm for operating ride-pooling services with en-route transfers.

2. Identification and integration of a previously overlooked dimension in transfer decisions, leading to a more comprehensive cost function approximation model.

3. Conducting a rigorous simulation-based experiment, utilizing real-world data to compare various operational strategies and validate the effectiveness of the proposed algorithm.

As cities continue to grapple with the challenges of urban mobility, studies like this offer a beacon of hope, paving the way for more efficient, sustainable, and accessible transportation systems.

This research was supported by MOIA.

Namdarpour, F., Liu, B., Kuehnel, N., Zwick, F., & Chow, J. Y. J. (2024). On non-myopic internal transfers in large-scale ride-pooling systems. Transportation Research Part C: Emerging Technologies, 162, 104597. https://doi.org/10.1016/j.trc.2024.104597

Facial recognition technology has made great strides in accuracy thanks to advanced artificial intelligence (AI) models trained on massive datasets of face images.

These datasets often lack diversity in terms of race, ethnicity, gender, and other demographic categories, however, causing facial recognition systems to perform worse on underrepresented demographic groups compared to groups ubiquitous in the training data. In other words, the systems are less likely to accurately match different images depicting the same person if that person belongs to a group that was insufficiently represented in the training data.

This systemic bias can jeopardize the integrity and fairness of facial recognition systems deployed for security purposes or to protect individual rights and civil liberties.

Researchers at NYU Tandon School of Engineering are tackling the problem. In a recent paper, a team led by Julian Togelius, Associate Professor of Computer Science and Engineering (CSE) revealed it successfully reduced facial recognition bias by generating highly diverse and balanced synthetic face datasets that can train facial recognition AI models to produce more fair results. The paper’s lead author is Anubhav Jain, Ph.D. candidate in CSE.

The team applied an "evolutionary algorithm" to control the output of StyleGAN2, an existing generative AI model that creates high-quality artificial face images and was initially trained on the Flickr Faces High Quality Dataset (FFHQ). The method is a "zero-shot" technique, meaning the researchers used the model as-is, without additional training.

The algorithm the researchers developed searches in the model’s latent space until it generates an equal balance of synthetic faces with appropriate demographic representations. The team was able to produce a dataset of 13.5 million unique synthetic face images, with 50,000 distinct digital identities for each of six major racial groups: White, Black, Indian, Asian, Hispanic and Middle Eastern.  

The researchers then pre-trained three facial recognition models — ArcFace, AdaFace and ElasticFace — on the large, balanced synthetic dataset they generated.

The result not only boosted overall accuracy compared to models trained on existing imbalanced datasets, but also significantly reduced demographic bias. The trained models showed more equitable accuracy across all racial groups compared to existing models exhibiting poor performance on underrepresented minorities.

The synthetic data proved similarly effective for improving the fairness of algorithms analyzing face images for attributes like gender and ethnicity categorization.

By avoiding the need to collect and store real people's face data, the synthetic approach delivers the added benefit of protecting individual privacy, a concern when training AI models on images of actual people’s faces. And by generating balanced representations across demographic groups, it overcomes the bias limitations of existing face datasets and models.

The researchers have open-sourced their code to enable others to reproduce and build upon their work developing unbiased, high-accuracy facial recognition and analysis capabilities. This could pave the way for deploying the technology more responsibly across security, law enforcement and other sensitive applications where fairness is paramount.

This study — whose authors also include Rishit Dholakia (’22) MS in Computer Science, NYU Courant; and Nasir Memon, Dean of Engineering at NYU Shanghai, NYU Tandon ECE professor and faculty member of NYU Center for Cybersecurity — builds upon a paper the researchers shared at the IEEE International Joint Conference on Biometrics (IJCB), September 25-28, 2023.

Anubhav Jain , Rishit Dholakia , Nasir Memon , et al. Zero-shot demographically unbiased image generation from an existing biased StyleGAN. TechRxiv. December 02, 2023

In the clandestine world of biochemical warfare, researchers are continuously seeking innovative strategies to counteract lethal agents. Researchers led by Jin Kim Montclare, Professor in the Department of Chemical and Biomolecular Engineering, have embarked on a pioneering mission to develop enzymatic defenses against chemical threats, as revealed in a recent study.

The team's focus lies in crafting enzymes capable of neutralizing notorious warfare agents such as VX, renowned for their swift and devastating effects on the nervous system. Through meticulous computational design, they harnessed the power of enzymes like phosphotesterase (PTE), traditionally adept at detoxifying organophosphates found in pesticides, to target VX agents.

The study utilized computational techniques to design a diverse library of PTE variants optimized for targeting lethal organophosphorus nerve agents. Leveraging advanced modeling software, such as Rosetta, the researchers meticulously crafted enzyme variants tailored to enhance efficacy against these formidable threats. When they tested these new enzyme versions in the lab, they found that three of them were much better at breaking down VX and VR. Their findings showcased the effectiveness of these engineered enzymes in neutralizing these chemicals.

A key problem in treating these agents lies in the urgency of application. In the event of exposure, rapid intervention becomes paramount. The research emphasizes potential applications, ranging from prophylactic measures to immediate administration upon exposure, underscoring the imperative for swift action to mitigate the agents' lethal effects.

Another key issue is protein stability — ensuring that the proteins can stay intact and at the site of affected tissue which is crucial for therapeutic applications. Ensuring enzymes remain stable within the body enhances their longevity and effectiveness, offering prolonged protection against chemical agents.

Looking ahead, Montclare's team aims to optimize enzyme stability and efficacy further, paving the way for practical applications in chemical defense and therapeutics. Their work represents a beacon of hope in the ongoing battle against chemical threats, promising safer and more effective strategies to safeguard lives.

Kronenberg, J., Chu, S., Olsen, A., Britton, D., Halvorsen, L., Guo, S., Lakshmi, A., Chen, J., Kulapurathazhe, M. J., Baker, C. A., Wadsworth, B. C., Van Acker, C. J., Lehman, J. G., Otto, T. C., Renfrew, P. D., Bonneau, R., & Montclare, J. K. (2024). Computational design of phosphotriesterase improves v‐agent degradation efficiency. ChemistryOpen. https://doi.org/10.1002/open.202300263

Human brain states are unobserved states that can constantly change due to internal and external factors, including cognitive arousal, a.k.a. intensity of emotion, and cognitive performance states. Maintaining a proper level of cognitive arousal may result in being more productive throughout daily cognitive activities. Therefore, monitoring and regulating one’s arousal state based on cognitive performance via simple everyday interventions such as music is a critical topic to be investigated.

Researchers from NYU Tandon led by Rose Faghih — inspired by the Yerkes-Dodson law in psychology, known as the inverted-U law — investigated the arousal-performance link throughout a cognitive task in the presence of personalized music. The Yerkes-Dodson law states that performance is a function of arousal and has an inverted-U shaped relationship with cognitive arousal, i.e., a moderate level of arousal results in optimal performance, on the other hand, an excessively high level of arousal may result in anxiety, while a deficient level of arousal may be followed by boredom.

In this study, participants selected music with calming and exciting music components to mimic the low and high-arousing environment. To decode the underlying arousal and performance with respect to everyday life settings, they used peripheral physiological data as well as behavioral signals within the Bayesian Decoders. In particular, electrodermal activity (EDA) has been widely used as a quantitative arousal index. In parallel, behavioral data such as a sequence of correct/incorrect responses and reaction time are common cognitive performance observations.

The decoded arousal and performance data points in the arousal-performance frame depict an inverted U shape, which conforms with the Yerkes-Dodson law. Also, findings present the overall better performance of participants within the exciting background music. Considering the Yerkes-Dodson law, we develop a performance-based arousal decoder that can preserve and account for the cognitive performance dynamic. Such a decoder can provide a profound insight into how physiological responses and cognitive states interplay to influence productivity.

Although several factors, such as the nature of the cognitive task, the participant’s baseline, and the type of applied music, can impact the outcome, it might be feasible to enhance cognitive performance and shift one’s arousal from either the left or right side of the curve using music. In particular, the baseline of arousal level varies among humans, and the music may be selected to set the arousal within the desired range. The outcome of this research can advance researchers closer to developing a practical and personalized closed-loop brain-computer interface for regulating internal brain states within everyday life activities.

S. Khazaei, M. R. Amin, M. Tahir and R. T. Faghih, "Bayesian Inference of Hidden Cognitive Performance and Arousal States in Presence of Music," in IEEE Open Journal of Engineering in Medicine and Biology, doi: 10.1109/OJEMB.2024.3377923.

The human body can be viewed as a complex collection of interrelated control systems. All these systems quietly function to maintain different variables or states within our bodies. Now the states themselves aren’t always easily accessible, but the physiological changes and signals they give rise to, in a number of cases, are. Hence, one can view the human body through the lens of control theory, and this allows the mathematical tools typically used in control engineering to be applied to understanding physiology.

This is indeed the case with certain states such as emotion, cognition and energy — the values themselves aren’t easily accessible, but the changes in sweat secretions, heart rate and hormone secretions can indeed be measured. Consequently, we can use tools from control systems to estimate these unknown quantities.

In several cases, the observations to which the underlying state variables are related to are “spikey” or pulsatile in appearance and nature. Skin conductance is one such example. Bursts of neural activity to the sweat glands are responsible for the spiky appearance of a skin conductance signal. Likewise, the hormone cortisol — the body’s main stress hormone — is also secreted in pulses. Moreover, both cortisol and skin conductance signals can be modeled in strikingly similar ways mathematically. In the case of skin conductance and cortisol however, the signals aren’t purely spikey or pulsatile in nature, but are also accompanied by or incorporate a continuous-valued baseline or biochemical concentration level.

Previous work attempting to estimate underlying aspects of emotion from skin conductance or energy levels from cortisol have not considered this specific formulation of the observations. In our present work, we consider modeling the observations from skin conductance and cortisol as a marked point process (MPP) coupled together with a continuous-valued variable. This matches the inherent constituent components of the signals much better.

Researchers at NYU Tandon led by Rose Faghih, Associate Professor of Biomedical Engineering, developed a decoder to estimate an aspect of emotion known as sympathetic arousal from skin conductance and an energy state from cortisol measurements. The model was able to capture more subtle, fine-grained variations in the underlying states compared to estimates from some of our earlier models. Cortisol has a characteristic daytime vs. nighttime secretory pattern and their model was able to estimate energy levels consistent with these expectations. Estimates of sympathetic arousal based on skin conductance were also high during certain stressors and lower during relaxation.

They also obtained sympathetic arousal estimates that were consistent with blood flow patterns in the brain in an additional experiment as well. Here, blood flow was measured using a technique known as functional Near Infrared Spectroscopy (fNIRS). The algorithm l they developed on this occasion also had a superior capability with dealing with what is known as overfitting in comparison to our earlier methods.

D. S. Wickramasuriya, S. Khazaei, R. Kiani and R. T. Faghih, "A Bayesian Filtering Approach for Tracking Sympathetic Arousal and Cortisol-Related Energy From Marked Point Process and Continuous-Valued Observations," in IEEE Access, vol. 11, pp. 137204-137247, 2023, doi: 10.1109/ACCESS.2023.3334974

Data science methods can help overcome challenges in measuring and analyzing social determinants of health (SDoH), according to a paper published in Lancet Digital Health, helping mitigate the root causes of health inequities that are not fully addressed through health care spending or lifestyle choices.

The paper came out of the NYU-Moi Data Science Social Determinants Training Program (DSSD), a collaboration between New York University, the NYU Grossman School of Medicine, Moi University, and Brown University that is funded by the National Institutes of Health (NIH). Through interdisciplinary training at NYU, DSSD aims to build a cohort of data science trainees from Kenya.

Rumi Chunara, associate professor at both NYU Tandon School of Engineering and NYU School of Global Public Health, is a DSSD Program Principal Investigator and wrote the paper with colleagues from DSSD’s collaborating institutions and the NIH.

SDoH are the diverse conditions in people's environments that affect their health, such as racism and climate. These conditions can negatively impact quality of life and health outcomes by shaping economic policies, social norms, and other environmental factors that consequently influence individual behaviors.

According to the researchers, the three main challenges — and potential solutions — in studying SDoH are:

1. SDoH data is hard to measure, especially at multiple levels like individual, community, and national, with racism being one notable example. Data science methods can help capture social determinants of health not easily quantified, like racism or climate impacts, from unstructured data sources including social media, notes, or imagery. For example, natural language processing can extract housing insecurity from medical notes, and deep learning can parse environmental factors from satellite imagery. These unstructured sources provide diverse insights compared to tabular, structured data, but also may contain biases requiring careful inspection. Incorporating social determinants from flexible, unstructured sources into analyses can better capture the heterogeneity of health effects across different populations.
2. SDoH impact health through complex, nonlinear pathways over time. Social factors like income or education are farther removed from health outcomes than medical factors. They affect health through complicated chains of intermediate factors that can also flow back to influence the social factors. For instance, income provides resources for healthy behaviors that improve health, while poor health hinders income. Advanced modeling techniques like machine learning can handle these tangled relationships between many variables better than simpler statistical models. Models that simulate individuals' behaviors and interactions allow studying how health patterns emerge from social factors. This captures the real-world complexity traditional models may miss between broad social conditions and individual health.
3. It takes a long time, sometimes decades, to observe how SDoH ultimately affect health outcomes. For example, lack of fresh produce and recreation options leads to poor nutrition, but chronic diseases take decades to develop. Longitudinal data over such time spans is rare, especially globally. Collecting representative surveys is resource-intensive. But novel digital data like mobile usage, purchases, or satellite imagery can provide longitudinal views at granular place and time scales. With proper privacy protections and population considerations, these new data managed with data science methods can help model social determinants' long-term health impacts.

Fully leveraging data science for SDoH research requires diverse experts working collaboratively across disciplines, according to the researchers. Training more data scientists, especially from underrepresented backgrounds, in SDoH is pivotal. Developing local data science skills grounded in community knowledge and values is also vital.

Along with Chunara, the paper’s authors are: Jessica Gjonaj from NYU School of Global Public Health and NYU Grossman; Rajesh Vedanthan from NYU Grossman; Eileen Immaculate, Iris Wanga, Judith Mangeni and Ann Mwangi from the College of Health Sciences at Moi University (Eldoret, Kenya); James Alaro and Lori A. J. Scott-Sheldon from the National Institutes of Health; and Joseph Hogan from Brown University.

Chunara, R., Gjonaj, J., Immaculate, E. et al. Social determinants of health: the need for data science methods and capacity. The Lancet Digital Health, 6(4), e235–e237 (2024). https://doi.org/10.1016/S2589-7500(24)00022-0

In the wake of the COVID-19 pandemic, scientists have been racing to develop effective treatments and preventatives against the virus. A recent scientific breakthrough has emerged from the work of researchers aiming to combat SARS-CoV-2, the virus responsible for COVID-19.

Led by Jin Kim Montclare and her team, the study focuses on the design and development of a novel protein capable of binding to the spike proteins found on the surface of the coronavirus. The goal behind this innovative approach is twofold: first, to identify and recognize the virus for diagnostic purposes, and second, to hinder its ability to infect human cells.

The engineered protein, resembling a structure with five arms, exhibits a unique feature—a hydrophobic pore within its coiled-coil configuration. This feature enables the protein not only to bind to the virus but also to capture small molecules, such as the antiviral drug Ritonavir.

Ritonavir, already utilized in the treatment of SARS-CoV-2 infections, serves as a logical choice for integration into this protein-based therapeutic. By incorporating Ritonavir into the protein, the researchers aim to enhance the treatment's efficacy while simultaneously targeting the virus directly.

The study marks a significant advancement in the fight against COVID-19, showcasing a multifaceted approach to combating the virus. Through a combination of protein engineering and computational design, the team has devised a promising strategy that may revolutionize current treatment modalities.

Although the research is still in its early stages, with no human or animal trials conducted as yet, the findings offer a proof of principle for the therapeutic potential of the designed protein. The team has demonstrated its ability to enhance the protein's binding affinity to the virus spike protein, laying the groundwork for future investigations.

The potential applications of this protein-based therapeutic extend beyond COVID-19. Its versatility opens doors to combating a range of viral infections, offering a dual mode of action—preventing viral entry into human cells and neutralizing virus particles.

Furthermore, the success of this study underscores the importance of computational approaches in protein design. By leveraging computational tools such as Rosetta, the researchers have accelerated the process of protein engineering, enabling rapid iterations and optimization.

The development of this novel protein represents a significant step forward in the ongoing battle against COVID-19. As research progresses, the integration of computational design and protein engineering holds promise for the development of innovative therapeutics with broad-spectrum antiviral capabilities. While challenges remain, this study offers hope for a future where effective treatments against emerging viral threats are within reach.