Denisa Qori McDonald
Machine Learning Researcher and Engineer building reliable learning systems for structured sequential data
I design machine learning systems for multivariate, human-generated data, focusing on preserving useful signal under noise, sparsity, limited observability, and distribution shift.
My work spans knitted textile sensors, conversational behavior, and physiological time series. While these domains are distinct, they raise the same underlying machine learning questions: how to learn from sparse observations, generalize across users and conditions, and build representations that remain resilient to uncontrolled conditions.
I also build reusable ML infrastructure: modular pipelines for preprocessing, representation learning, evaluation, and deployment under variable real-world conditions. These questions are central to modern ML systems, including foundation models, where representation quality, robustness under shift, and behavior beyond controlled environments are central challenges.
What I Work On
Robust representation learning
Designing features and learned representations for sequential and multivariate data that remain informative across users, environments, and naturalistic variation. I focus on recovering stable information from noisy, sparse, or partially observed signals.
Temporal modeling under uncontrolled conditions
Developing machine learning approaches for behavioral, physiological, and sensor-based time series, combining signal processing, feature construction, and learned models to capture meaningful temporal patterns across users, settings, and conditions.
End-to-end ML systems
Building reusable machine learning pipelines from multichannel inputs and preprocessing through modeling, evaluation, and deployment. I see architecture, evaluation design, and implementation constraints as part of the ML problem itself, because they shape what a model can learn, how it can be tested, and whether it can be used reliably.
Research Perspective
The question running through my work is how learning systems remain reliable when data are noisy, users differ, and practical conditions shift.
Across projects, I have considered high predictive performance as one measure of success, but not the only one. I also investigate whether the learned structure is valid beyond the setting in which it was first evaluated, which naturally leads to studying representation quality, robustness, and evaluation under cross-context variation.
Selected Work
Representation Learning for Sparse Textile Sensing
Developed machine learning methods for touch localization and gesture recognition from minimally wired knitted sensors, showing how stable representations and temporal models can recover human intent under hardware, physical, and interaction constraints.
This work frames sparse textile sensing as a machine learning problem in limited observability: learning useful representations from minimal, noisy measurements while testing whether they generalize across users and conditions.
Computational Modeling of Naturalistic Social Behavior
Built interpretable temporal representations of head movement and social behavior in face-to-face interaction, enabling quantitative modeling of clinically relevant behavioral signals in ecologically valid settings.
This project shows how noisy temporal behavior can be modeled as structured signal rather than incidental variation, while preserving interpretability for developmental and clinical questions.
Physiological Time-Series Modeling
Developed early machine learning pipelines for physiological signals, including fNIRS, establishing foundations in signal representation, temporal modeling, and sound analysis of noisy human-generated data.
This work helped me standardize my approach to real-world machine learning, grounding it not only in representation design, but also in preprocessing, repeatability, replication, and cross-condition evaluation.