Sitemap
A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.
Pages
Posts
Nutrition
Published:
LET’S TALK ABOUT NUTRITION
Endurance? What is it?
The Science Behind Running - Lactate
Published:
The Science Behind Running - Lactate
“Lactate threshold” is a critical yet often misunderstood term in the field of endurance training. This blog seeks to clarify its role, its importance for athletes—especially runners—and the methods that can be employed to optimize this vital physiological indicator for improved running performance.
The Science Behind Running - VO2
Published:
The Science Behind Running - VO2
The Benefits of rotating running shoes
Published:
THE BENEFITS OF ROTATING RUNNING SHOES: ENHANCING PERFORMANCE AND PREVENTING INJURIES
Unraveling the mystery of running shoes carbon plated running shoes part II
Published:
When you dive into the vast ocean of running shoes, a term you’ll likely come across is “drop.” If you’re not a seasoned runner, this term might be as mysterious as the dark side of the moon. In its simplest form, the “drop” in a running shoe, also known as the “heel-to-toe drop,” is the difference in cushioning height between the heel and the toe of the shoe, usually measured in millimeters (mm).
Unraveling the mystery of running shoes carbon plated running shoes part I
Published:
In today’s world, running shoes boast a combination of innovative features tailored to each runner’s unique needs [2]. Think shoe weight and longitudinal stiffness for enhancing running economy (RE), diverse midsole hardness and upper construction for optimal comfort, or slimmer midsoles and reduced heel-toe drops for injury prevention [2].
Understanding Running Dynamics Parameters
Published:
Nowadays, many smart watches record and measure many parameters other than heart rate, such as cadence, vertical ratio, vertical oscillation, and many more. Runners can gain much better insight into their techniques and training progress by using these measurements.
portfolio
publications
Matched Effective Sample Size for Temporal Evaluation with Conservative TCN Bounds
Under review
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, show in a 640-run controlled sweep that fixed-length comparisons can reverse the apparent ordering, and pair it with a conservative generalization bound for temporal convolutional networks on beta-mixing sequences.
Differentiable Randers-Finsler Eikonal Solvers
Computer Graphics Forum (Proceedings of SGP 2026)
Direction-dependent (Randers-Finsler) distance fields describe wavefront phenomena such as wildfire spread, but classical Eikonal solvers are not differentiable, so their metrics cannot be learned from data. We make a GPU-friendly fast-sweeping solver exactly differentiable through implicit differentiation, recover spatially varying anisotropic metrics and drift fields from sparse arrival times, and learn data-driven spread models on real wildfire perimeters.
The Residual Stream's Effective Depth
Asian Conference on Machine Learning (ACML 2026), PMLR
Adding layers does not guarantee that a language model performs more distinct computation, and there was no calibrated way to measure how much of its depth is actually used. Effective depth treats the residual stream as a layer-indexed process and summarizes how quickly representations decorrelate: fifteen of sixteen decoder-only models (Qwen3.5, OLMo-2, Pythia) fall below the closed-form orthogonal-update reference, and matched controls attribute the gap to correlated residual updates rather than unused depth.
The Shuffle Gap: Diagnosing Personalization Collapse in LLMs
Under review
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.
Invertible Query-Key Coupling Composes with Attention Mechanisms
Asian Conference on Machine Learning (ACML 2026), PMLR
Most attention variants change how scores are computed; this work instead transforms each token's query and key jointly before scoring, through an invertible, identity-initialized coupling that adds about 0.3% parameters and leaves the softmax untouched. Because it acts before scoring it composes with existing mechanisms: on WikiText-103 it yields statistically significant perplexity gains on top of Differential Attention at both 150M and 455M parameters.
Predicting Cardiovascular Drift from 3D Running Pose via Spatio-Temporal Graph Attention
Under review
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.
Using Deep Reinforcement Learning for mmWave Real-Time Scheduling
14th International Conference on Network of the Future (NoF), 2023
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.
Data-Driven Cellular Network Selector for Vehicle Teleoperations
15th International Conference on Network of the Future (NoF), 2024
Teleoperated vehicles stream video over cellular links whose packet loss and latency change from second to second. The Active Network Selector (ANS), a time-series model trained on 100 real test drives, decides for every video packet which of several cellular networks to send it over and beats the commercial baseline on both loss and latency.
Estimating the Number of HTTP/3 Responses in QUIC Using Deep Learning
IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN), 2025
QUIC encrypts almost everything, so network middleboxes can no longer see how many HTTP/3 responses a connection carries, a quantity that matters for load balancing and flood-attack detection. We turn encrypted QUIC traces into image sequences and train a deep model with a tailored loss, reaching up to 97% accuracy on more than 7 million images from 100,000 traces across 44,000 websites.
WearableMil: An End-to-End Framework for Military Activity Recognition and Performance Monitoring
IEEE 13th International Conference on Healthcare Informatics (ICHI), 2025
Musculoskeletal injuries during military training are common, yet field activity monitoring has no predefined sessions and plenty of missing data. From six months of Garmin smartwatch data on 135 soldiers, a hierarchical deep learning framework recognizes activities at 93.8% accuracy (83.8% cross-user), and physiologically informed imputation cuts unknown sleep states from 40% to 3.7%.
Exploring QUIC Dynamics: A Large-Scale Dataset for Encrypted Traffic Analysis
IEEE International Mediterranean Conference on Communications and Networking (MeditCom 2025)
Public QUIC datasets are outdated, small, or stripped of the SSL keys needed to label encrypted traffic. VisQUIC releases over 100,000 labeled QUIC traces with keys from more than 40,000 websites, together with image representations that let models estimate HTTP/3 request-response counts from encrypted traffic alone at up to 92% accuracy.
ML-ENHANCE: A Machine Learning Framework for Hospital Readmission Prediction with Cost-Optimized Decision Thresholds
Under review
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.
From Lab to Wrist: Bridging Metabolic Monitoring and Consumer Wearables for Heart Rate and Oxygen Consumption Modeling
27th ACM International Conference on Multimodal Interaction (ICMI 2025), oral presentation
Instantaneous oxygen consumption (VO₂) is the reference measure of metabolic demand, but it normally requires a lab-grade breath analyzer. This framework predicts full VO₂ trajectories from smartwatch and chest-strap data alone: a physiologically constrained ODE with a neural Kalman filter models heart rate at 2.81 bpm MAE on over 3 million observations, and a sequence-to-sequence VO₂ model needs only one second of calibration to reach about 11% MAPE.
talks
Understanding running dynamics parameters
Published:
Nowadays, many smart watches record and measure many parameters other than heart rate, such as cadence, vertical ratio, vertical oscillation, and many more. Runners can gain much better insight into their techniques and training progress by using these measurements.