Publications
Publications
Peer-reviewed papers first, then preprints and papers under review, each with a short summary. Also on Google Scholar.
Peer-reviewed
Invertible Query-Key Coupling Composes with Attention Mechanisms
An identity-initialized, invertible coupling applied before scoring that adds 0.3% parameters and composes with Differential Attention, with significant gains at 150M and 455M parameters.
The Residual Stream's Effective Depth
A calibrated scalar for how much of a transformer's depth is distinct computation: fifteen of sixteen decoder-only models fall below the orthogonal-update reference, and matched controls trace the gap to correlated residual updates.
Differentiable Randers-Finsler Eikonal Solvers
A GPU fast-sweeping solver made exactly differentiable through implicit differentiation, used to recover anisotropic metrics from sparse arrival times and learn spread models on real wildfire perimeters.
From Lab to Wrist: Bridging Metabolic Monitoring and Consumer Wearables for Heart Rate and Oxygen Consumption Modeling
Full VO₂ trajectories from smartwatch and chest-strap data alone: a physiologically constrained ODE with a neural Kalman filter for heart rate, and a one-second-calibration VO₂ model at about 11% MAPE.
Exploring QUIC Dynamics: A Large-Scale Dataset for Encrypted Traffic Analysis
A public dataset of over 100,000 labeled QUIC traces from more than 40,000 websites, rendered as image sequences so encrypted traffic can be studied with standard vision models.
WearableMil: An End-to-End Framework for Military Activity Recognition and Performance Monitoring
Hierarchical activity recognition for 135 soldiers over six months (15M+ data-minutes); physiologically informed imputation cuts unknown states from 40% to 3.6%.
Estimating the Number of HTTP/3 Responses in QUIC Using Deep Learning
DecQUIC turns encrypted QUIC traces into image sequences and estimates how many HTTP/3 responses a passive observer saw, reaching up to 97% accuracy on known and unseen servers.
Data-Driven Cellular Network Selector for Vehicle Teleoperations
A CNN-LSTM Active Network Selector trained on 100 cellular test drives picks the network for each video packet in real time and beats the commercial baseline on both loss and latency.
Using Deep Reinforcement Learning for mmWave Real-Time Scheduling
Real-time link scheduling in 5G millimeter-wave mesh networks has to decide which links to activate, and at what power, within a few milliseconds per slot. AARL is a model-free deep reinforcement learning scheduler that makes these decisions inside the strict time constraints of 5G mmWave networks.
Preprints and papers under review
- Paper
Predicting Cardiovascular Drift from 3D Running Pose via Spatio-Temporal Graph Attention
Cardiovascular drift, the gradual rise in heart rate at a constant running effort, is a key marker of endurance load but normally needs a wearable to measure. A spatio-temporal graph-attention model infers it directly from monocular 3D running pose, linking biomechanics to cardiovascular load with no body-worn sensor and giving a physiologically interpretable estimator for endurance monitoring.
The Shuffle Gap: Diagnosing Personalization Collapse in LLMs
Personalized language models are usually judged by held-out perplexity after conditioning on a user's history, which cannot tell whether the model used that user's data or merely learned a population-level adapter. The shuffle gap, the change in perplexity when each user's support set is swapped for another user's, exposes cases where large aggregate gains coexist with near-zero dependence on the matched user, making it a necessary audit for perplexity-based personalization claims.
Evaluating Temporal Models at Matched Information
Consecutive samples in a time series are correlated, so raw sequence length overstates how much information a temporal benchmark contains. We propose an evaluation protocol that matches an operational effective sample size across regimes and show, across more than 2,700 controlled runs with TCN, LSTM and causal Transformer predictors, that stronger dependence does not significantly increase train-test gaps once information is matched; a conservative generalization bound for temporal convolutional networks on beta-mixing sequences completes the protocol.
ML-ENHANCE: A Machine Learning Framework for Hospital Readmission Prediction with Cost-Optimized Decision Thresholds
Thirty-day hospital readmissions affect 15 to 20% of patients and cost over 40 billion dollars a year, yet common risk scores ignore the cost of acting on a prediction. Trained on 113,312 internal-medicine admissions, ML-ENHANCE reaches AUC 0.752 against 0.696 for the clinical ENHANCE score and 0.676 for HOSPITAL, and its cost-optimized thresholds avoid thousands of unnecessary interventions per year at matched sensitivity.