About

Benjamin Marlin is an associate professor in the College of Information and Computer Sciences at the University of Massachusetts Amherst. He serves as the director of the REML lab, where his work involves leading research efforts in machine learning and its applications. The lab includes a diverse group of researchers and students working on topics such as deep learning for multi-modal detection, Bayesian time series modeling, reinforcement learning, and modeling incomplete wearable sensor data using deep time series models. Marlin's research group has a strong focus on computationally efficient Bayesian deep learning methods, structured prediction, domain adaptation, active learning, and the analysis of irregularly sampled time series data, particularly with applications in mobile health and behavioral data analysis. His mentorship has guided numerous PhD and MS students, as well as undergraduate researchers, contributing to advancements in machine learning methodologies and their practical deployment in health-related sensor systems.

Research topics

• Computer Science
• Data science
• Artificial Intelligence
• Medicine
• Embedded system
• Nursing
• Telecommunications
• Operating system
• Computer architecture
• World Wide Web
• Psychology
• Human–computer interaction
• Social psychology
• Applied psychology

Selected publications

• Secondary Analysis to Advance Characterization of On-body Electrocardiographic Sensors in a Clinical Cocaine Self-administration Paradigm

  Journal of Addiction Medicine · 2025-04-03 · 1 citations

  articleOpen access

  OBJECTIVES: Our group has previously established how remote on-body electrocardiogram (ECG) sensors may discriminate cocaine use from other sympathomimetic conditions. The current analyses assess whether discriminatory power is mainly driven by differences in heart rate between conditions. METHODS: Individuals who use cocaine (N = 11) wore ECG sensors during (1) cocaine self-administration, (2) methylphenidate administration, (3) aerobic exercise, and (4) tobacco use (N = 9). Primary outcomes included: (1) time elapsed between 2 successive R waves (ie, RR interval), (2) ECG interval proxies, and (3) waveforms. ECG traces were matched for heart rate between conditions for all discriminations. RESULTS: ECG interval proxies and waveforms exhibited high discriminatory power in distinguishing cocaine use from methylphenidate, exercise, and tobacco use, with mean areas under the receiver operating characteristics ranging from 0.87 to 0.99, while RR-related measures ranged from 0.49 to 0.5, reflecting low discriminatory power. CONCLUSION: Our results suggest that the ECG sensors' discriminatory power is based on nuances in ECG data beyond mere changes in heart rate.

  PublisherOA PDFDOI

• ACE and Diverse Generalization via Selective Disagreement

  ArXiv.org · 2025-09-09

  preprintOpen access

  Deep neural networks are notoriously sensitive to spurious correlations - where a model learns a shortcut that fails out-of-distribution. Existing work on spurious correlations has often focused on incomplete correlations,leveraging access to labeled instances that break the correlation. But in cases where the spurious correlations are complete, the correct generalization is fundamentally \textit{underspecified}. To resolve this underspecification, we propose learning a set of concepts that are consistent with training data but make distinct predictions on a subset of novel unlabeled inputs. Using a self-training approach that encourages \textit{confident} and \textit{selective} disagreement, our method ACE matches or outperforms existing methods on a suite of complete-spurious correlation benchmarks, while remaining robust to incomplete spurious correlations. ACE is also more configurable than prior approaches, allowing for straight-forward encoding of prior knowledge and principled unsupervised model selection. In an early application to language-model alignment, we find that ACE achieves competitive performance on the measurement tampering detection benchmark \textit{without} access to untrusted measurements. While still subject to important limitations, ACE represents significant progress towards overcoming underspecification.

  PublisherOA PDFDOI

• Using LLMs to improve RL policies in personalized health adaptive interventions

  2025-01-01 · 2 citations

  articleOpen accessSenior author

  Reinforcement learning (RL) is increasingly used in the healthcare domain, particularly for the development of personalized adaptive health interventions.However, RL methods are often applied to this domain using small state spaces to mitigate data scarcity.In this paper, we aim to use Large Language Models (LLMs) to incorporate text-based user preferences and constraints, to update the RL policy.The LLM acts as a filter in the action selection.To evaluate our method, we develop a novel simulation environment that generates text-based user preferences and incorporates corresponding constraints that impact behavioral dynamics.We show that our method can take into account the text-based user preferences, while improving the RL policy, thus improving personalization in adaptive intervention.

  PublisherOA PDFDOI

• Practical Considerations for Failure Resilient ML Systems at the Edge

  2025-10-06

  article

  Machine learning at the edge is increasingly embedded in our daily lives, supporting applications running on smartphones, wearables, and industrial IoT. Prior work has mainly focused on resource efficiency and latency optimization through innovations in compact model design and resource-management techniques to meet stringent performance targets. However, edge devices and networks are inherently subject to failures and performance fluctuations, requiring an emphasis on failure resilience, especially in resource-constrained edge environments. Although recent studies have proposed resource-aware mechanisms and failure-aware models to improve the resilience of machine learning systems at the edge, many overlook deployment overheads that impede the adoption of these approaches. In this paper, we highlight practical considerations that affect failure-detection and recovery times and analyze how these considerations shape system design. We outline future research directions to enable practical, failure-resilient machine-learning systems at the edge.

  PublisherDOI

• End-to-End Differentiable Multi-View Tracking: Architecture and Fine-Tuning Experiments

  2025-07-07 · 1 citations

  articleSenior author

  In this work, we develop an end-to-end differentiable multi-view visual tracking architecture and explore fine-tuning model parameters via gradient-based optimization and automatic differentiation. We consider a setting with multiple camera nodes distributed in the tracking environment that collaboratively track objects. The architecture that we construct includes within-image-plane deep learning-based detection models, probabilistic camera models, object dynamics models, and an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$N$</tex>-object Kalman filter-based tracking model. We demonstrate fully differentiable choices for each of these components, enabling learning and fine-tuning of the parameters of all system components based on different forms of supervision. Our results show performance gains for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$N$</tex>-object tracking when fine-tuning the parameters of the system for end-to-end tracking performance.

  PublisherDOI

• Enhancing Adaptive Behavioral Interventions with LLM Inference from Participant-Described States

  ArXiv.org · 2025-07-05

  articleOpen accessSenior author

  The use of reinforcement learning (RL) methods to support health behavior change via personalized and just-in-time adaptive interventions is of significant interest to health and behavioral science researchers focused on problems such as smoking cessation support and physical activity promotion. However, RL methods are often applied to these domains using a small collection of context variables to mitigate the significant data scarcity issues that arise from practical limitations on the design of adaptive intervention trials. In this paper, we explore an approach to significantly expanding the state space of an adaptive intervention without impacting data efficiency. The proposed approach enables intervention participants to provide natural language descriptions of aspects of their current state. It then leverages inference with pre-trained large language models (LLMs) to better align the policy of a base RL method with these state descriptions. To evaluate our method, we develop a novel physical activity intervention simulation environment that generates text-based state descriptions conditioned on latent state variables using an auxiliary LLM. We show that this approach has the potential to significantly improve the performance of online policy learning methods.

  PublisherOA PDF

• A dynamic Bayesian network approach to modeling engagement and walking behavior: insights from a yearlong micro-randomized trial ( <i>Heartsteps II</i> )

  Health Psychology and Behavioral Medicine · 2025-09-18

  articleOpen accessSenior author

  INTRODUCTION: Mobile health (mHealth) technologies such as wearable activity trackers (e.g. Fitbit) and digital applications (apps), can support behavior change in real-world contexts. Since effectiveness is dependent, in part, on participants' engagement with the digital technology (e.g. app page views) and the intervention components (e.g. anti-sedentary messages), there is a need for modeling approaches that support the investigation of engagement in digital interventions and the refinement of dynamic theories of behavior change. METHODS: Dynamic Bayesian Networks (DBN) were used to model the idiographic (individual) dynamic relationships between a participant's daily app engagement (page views), walking behavior, and intervention messages, accounting for context (e.g. temperature), and psychological variables (e.g. perceived restedness and perceived busyness). Additionally, we explored differences in the resulting DBN models between participants of Hispanic/Latino and non-Hispanic/Latino White backgrounds. RESULTS: Data from 10 participants in the HeartSteps II study (n = 5 Hispanic/Latinos and n = 5 non-Hispanic/Latino Whites) was used. Across participants (100%, n = 10), there was a strong positive effect of the number of messages/prompts received on their daily app page views with a predicted increase range of 12.84 (12.19-13.57) to 25.84 (24.28-27.59) app page views per day per message received. Among the majority of Hispanic/Latino participants (n = 4/5, 80%), there was a strong positive relationship between daily app page views and walking behavior with predictions ranging from a mean of 6.70 (6.37-7.05) to 10.93 (10.14-11.78) steps per minute of Fitbit wear time per app page view. Both groups showed idiographic differences in the effects of temperature and perceived busyness on walking behavior. CONCLUSION: The results demonstrate the benefits of DBNs to model the daily-level idiographic behavioral dynamics of engagement in digital intervention studies. This approach can be leveraged to support the refinement of dynamic theories of behavior change and improving personalized mHealth intervention strategies.

  PublisherDOI

• SigmaScheduling: Uncertainty-Informed Scheduling of Decision Points for Intelligent Mobile Health Interventions

  ArXiv.org · 2025-07-14

  preprintOpen access

  Timely decision making is critical to the effectiveness of mobile health (mHealth) interventions. At predefined timepoints called "decision points," intelligent mHealth systems such as just-in-time adaptive interventions (JITAIs) estimate an individual's biobehavioral context from sensor or survey data and determine whether and how to intervene. For interventions targeting habitual behavior (e.g., oral hygiene), effectiveness often hinges on delivering support shortly before the target behavior is likely to occur. Current practice schedules decision points at a fixed interval (e.g., one hour) before user-provided behavior times, and the fixed interval is kept the same for all individuals. However, this one-size-fits-all approach performs poorly for individuals with irregular routines, often scheduling decision points after the target behavior has already occurred, rendering interventions ineffective. In this paper, we propose SigmaScheduling, a method to dynamically schedule decision points based on uncertainty in predicted behavior times. When behavior timing is more predictable, SigmaScheduling schedules decision points closer to the predicted behavior time; when timing is less certain, SigmaScheduling schedules decision points earlier, increasing the likelihood of timely intervention. We evaluated SigmaScheduling using real-world data from 68 participants in a 10-week trial of Oralytics, a JITAI designed to improve daily toothbrushing. SigmaScheduling increased the likelihood that decision points preceded brushing events in at least 70% of cases, preserving opportunities to intervene and impact behavior. Our results indicate that SigmaScheduling can advance precision mHealth, particularly for JITAIs targeting time-sensitive, habitual behaviors such as oral hygiene or dietary habits.

  PublisherOA PDFDOI

• Dynamic modeling and system identification of user engagement in mHealth interventions using a Bayesian approach for missing data imputation

  Control Engineering Practice · 2025-06-29

  article
  PublisherDOI

• GPU-GLMB: Assessing the Scalability of GPU-Accelerated Multi-Hypothesis Tracking

  2025-10-06

  articleSenior author

  Much recent research on multi-target tracking has focused on multi-hypothesis approaches leveraging random finite sets. Of particular interest are labeled random finite set methods that maintain temporally coherent labels for each object. While these methods enjoy important theoretical properties as closed-form solutions to the multi-target Bayes filter, the maintenance of multiple hypotheses under the standard measurement model is highly computationally expensive, even when hypothesis pruning approximations are applied. In this work, we focus on the Generalized Labeled Multi-Bernoulli (GLMB) filter as an example of this class of methods. We investigate a variant of the filter that allows multiple detections per object from the same sensor, a critical capability when deploying tracking in the context of distributed networks of machine learning-based virtual sensors. We show that this breaks the inter-detection dependencies in the filter updates of the standard GLMB filter, allowing updates with significantly improved parallel scalability and enabling efficient deployment on GPU hardware. We report the results of a preliminary analysis of a GPU-accelerated implementation of our proposed GLMB tracker, with a focus on run time scalability with respect to the number of objects and the maximum number of retained hypotheses.

  PublisherDOI

Recent grants

• SCH: INT: Collaborative Research: Enhancing Context-Awareness and Personalization for Intensively Adaptive Smoking Cessation Messaging Interventions

  NSF · $640k · 2017–2023

• CAREER: Machine Learning for Complex Health Data Analytics

  NSF · $537k · 2014–2019

• Operationalizing Behavioral Theory for mHealth: Dynamics, Context, and Personalization

  NIH · $2.1M · 2018–2023

Frequent coauthors

• Deepak Ganesan

  42 shared

• Kanav Saraf

  Samueli Institute

  36 shared

• Peter C. Coyte

  University of Toronto

  36 shared

• Aman Mahajan

  University of Pittsburgh

  36 shared

• Rui Fu

  36 shared

• Noah Jennings

  Old Dominion University

  36 shared

• Shubham Jain

  Indian Institute of Technology Bombay

  36 shared

• Michael H. Wasko

  Oregon Health & Science University

  36 shared

Labs

• REMLPI

Education

• Ph.D., machine learning

  University of Toronto

• Other

  University of British Columbia

Awards & honors

• Fellow of the Pacific Institute for the Mathematical Science…
• Fellow of the Killam Trusts at the University of British Col…