In an era of rapidly evolving misinformation and disinformation, AI plays a dual role—both a tool for analyzing harmful narratives and a challenge for content moderation. This talk presents our research on how AI helps in understanding political discourse and detecting disinformation, offering key insights into social media dynamics. By examining interactions of over 6,500 U.S. state legislators on Twitter and Facebook, we demonstrate how platform policies and affordances shape political engagement in distinct ways. These differences not only reinforce partisan divides but also raise critical questions about platform-specific incentives. Beyond analysis, we develop AI-driven methods to detect conspiratorial narratives and evaluate generative AI's effectiveness in fact-checking misinformation. Comparing AI-generated credibility assessments with human annotations, our research highlights both the promise and risks of automated verification. Additionally, through collaborations with NGOs, we have gained valuable experience in deploying AI-assisted tools to support election integrity and enhance fact-checking efforts. With AI-generated disinformation on the rise and industry shifts, including Meta's changes to fact-checking programs and X/Twitter's legal battles with watchdogs, the need to understand AI's evolving role in the information ecosystem is more critical than ever.

In this hands-on, 6-hour workshop, you'll learn to build and deploy a complete machine learning application on Amazon Web Services (AWS), a leading cloud platform. Topics include data preprocessing, feature engineering, experiment tracking, model deployment, and performance monitoring. By the end, you'll be equipped to create scalable, production-ready ML solutions beyond Scikit-learn and Jupyter Notebook.

Privacy-preserving mechanisms, particularly those satisfying differential privacy, have become critical in almost all data release exercises. However, there are no widely accepted privacy mechanisms for releasing dependent data, such as time series data. We will begin with a short survey of the existing privacy mechanisms for time series and discuss the need to go beyond what is available. Based on a desideratum that balances the nuances of traditional statistical time series analysis and the demands of modern data privacy mechanisms, a formal privacy-utility framework will be presented for time series data. Theoretical and operational characteristics of the proposed framework will be studied and possible future directions will be discussed.

Typically, we are taught a lot of programming in school. However, when we go to industry, does the kind of programming we know remain the same? What constitutes good code? Tune in to the talk for answers.

Probabilistic Graphical Models (PGMs) are used broadly across many fields to model the connectivity relationships between entities in a network. This workshop introduces Bayesian Networks, a special class of directed and acyclic PGMs. BNs are “expert systems” widely used for inference and prediction. Methods for parameter and structural learning will be discussed and implemented using the R programming language. Probabilistic reasoning will be described for making predictions within these networks. In the second part of the workshop, attendees will have a “hands-on” experience with network construction, inference, and visualization using the R programming language. Programming experience is not required. This workshop serves as a preview for a 1-credit course in the IAD’s ‘summer stackable’ graduate elective series. 

Dr. Bylsma will review her recent work focused on using innovative ambulatory physiological designs, including physiologically-triggered prompts to examine emotional processes in daily life transdiagnostically, including adults with distress disorders and autism spectrum disorders. She will also discuss her recent collaborations with experts in affective computing and machine learning to examine dyadic affective processes in the context of psychotherapy, as well as potential future applications leveraging machine learning to predict episodes of emotion dysregulation and suicide risk in daily life.

Sensing is a universal task in science and engineering. Downstream tasks from sensing include learning dynamical models, inferring full state estimates of a system (system identification), control decisions and forecasting. These tasks are exceptionally challenging to achieve with limited sensors, noisy measurements and corrupt or missing data. Existing techniques typically use current (static) sensor measurements to perform such tasks and require principled sensor placement or an abundance of randomly placed sensors. In contrast, we propose a SHallow REcurrent Decoder (SHRED) neural network structure which incorporates (i) a recurrent neural network (LSTM) to learn a latent representation of the temporal dynamics of the sensors and (ii) a shallow decoder that learns a mapping between this latent representation and the high-dimensional state space. By explicitly accounting for the time-history or trajectory of the sensor measurements, SHRED enables accurate reconstructions with far fewer sensors, outperforms existing techniques when more measurements are available and is agnostic towards sensor placement. In addition, a compressed representation of the high-dimensional state is directly obtained from sensor measurements, which provides an on-the-fly compression for modeling physical and engineering systems. Forecasting is also achieved from the sensor time-series data alone, producing an efficient paradigm for predicting temporal evolution with an exceptionally limited number of sensors. In the example cases explored, including turbulent flows, complex spatio-temporal dynamics can be characterized with exceedingly limited sensors that can be randomly placed with minimal loss of performance.

This workshop provides a comprehensive overview of modern time series analysis, equipping participants with both foundational knowledge and cutting-edge techniques. We begin with fundamental concepts, including handling temporal data, effective visualization, and essential baseline methods such as exponential smoothing and ARIMA modeling. Building on this foundation, we explore standard techniques like state-space models and spectral methods, ensuring participants develop a strong analytical toolkit.

The latter half of the workshop focuses on recent advancements in time series. We introduce neural networks, Long Short-Term Memory (LSTM) models, and transformer-based architectures, discussing their strengths, limitations, and practical applications. Participants will learn how to fit models, interpret results, and understand when deep learning provides an advantage over traditional approaches. Through hands-on exercises and real-world examples, attendees will gain the skills necessary to apply modern time series methods effectively in research and industry settings.

Emotion regulation is an important transdiagnostic predictor of well-being and psychological symptoms, as well as a target in many interventions. Although theory predicts that emotion regulation strategies will be most effective when appropriately matched to the specific demands of a given situation, little is known about which strategies are most effective for whom and in what context. As such, there is limited knowledge to draw from in preventative and intervention efforts targeting emotion regulation. Community adults (N = 374) completed eight daily surveys for 10 days, assessing emotional episodes, the strategies used to attempt to regulate them, and relevant situational variables and individual differences. Random forest models were constructed to identify the most important predictors of perceived emotion regulation effectiveness, as well as robust interactions among variables. Results will be discussed regarding clinical assessment and intervention, including recent innovations such as just-in-time adaptive interventions.

As large language models (LLMs) become more widely used to facilitate communication, understanding their potential use in interactive designs is crucial. This talk addresses two aspects of LLMs in communicative settings: (1) their capability to adopt diverse human perspectives in conversation and (2) their utility in serving as different types of assistants in conversation. The first explores the personalization of LLMs, examining their capacity to mimic human behavior and preferences. I show that many LLMs are surprisingly limited in their ability to simulate human dialogues and incorporate demographic nuances in their responses, challenging the efficacy of persona-inspired prompts and social simulation. In the second, I reframe LLMs as potential conversational enhancers, drawing from a large randomized controlled experiment that examines the efficacy of different AI-based content creation tools on social media. I show how these tools can diversify and improve participation in online discussions, though authors and readers diverge in their views of AI assistance. Together, this talk highlights the dual challenges and opportunities of leveraging LLMs in communication.

Although research has made significant progress, harmful alcohol use remains a significant public health concern. Many cases of alcohol use disorder go undetected and there is a significant treatment gap. Only a small percentage of those in need receive treatment. These detection and treatment gaps are more pronounced among minoritized populations. Narrowing these gaps requires scalable methods for detection and assessment, and dissemination of evidenced-based interventions that can reach underserved populations, not only with respect to harmful alcohol use, but also detection of early risk and protective factors.

Leveraging artificial intelligence (AI) to understand, prevent, and treat harmful alcohol use and alcohol use disorder can yield significant advancements in the field. AI broadly refers to building hardware and developing software and applications to perform functions that usually require human intelligence. A common characteristic of AI tools is the ability to analyze enormous amounts of data to detect patterns, make predictions, and solve problems. AI can perform tasks faster, more accurately, and at greater scale than humans and has the potential to increase the efficacy, reach, and efficiency of detection and intervention efforts, and enhance the quality of research. AI is increasingly being used in medicine and mental health. However, the field of addiction science has been slower to adopt AI tools. This is a missed opportunity.

This talk describes the development of a proposal for a dynamic, integrated, and sustainable center (The Center on Alcohol and Artificial Intelligence Research; CAAIR) that brings together a community of AI and alcohol scientists and other stakeholders round the theme of leveraging AI tools to understand and treat harmful alcohol use, bridging existing gaps across the AI and mental health/addiction disciplines. The proposed center illustrates some of the many ways that AI tools may be used to address public health problems such as addictive behaviors. Considerations of the ethical application of AI, community engagement, and health disparities will be discussed.