Research & Multi-Task Learning
How do modelling and measurement choices change the conclusions we draw? My work spans learning systems, statistical inference and scientific computing.
Gradient interference and transfer in multi-task learning
I compare two tasks trained together on the same inputs, using a shared trunk and separate heads, with matched single-task networks. The aim is to understand when sharing helps or hurts and whether transfer can be predicted before training.
Do public records document PFAS drinking-water response handoffs? A 51-jurisdiction audit
preprint. Gustavo Pedro Ricou. Audits whether public monitoring, assistance and project records document traceable PFAS response handoffs across 50 US states and the District of Columbia. A reproducible documentation audit, not a measure of inaction, programme performance or health effects.
Don't look now: paying for documented need punishes discovery in US lead-pipe funding
preprint. Gustavo Pedro Ricou. Kristin C. Epstein. Derives how resolving unknown lead pipes can reduce expected funding under the existing US allocation rule, then develops an audit-anchored alternative that removes this discovery penalty within the model. Reproducible allocation analysis, not a field-tested policy intervention.
One equation allocates playing time in competitive team sports: the rulebook sets the arithmetic and the league sets the habit
preprint. Gustavo Pedro Ricou. Decomposes playing-time allocation into rule constraints, withdrawal timing and competition behaviour using cross-sport records. Substitution chains and player re-entry limit the scope of a common identity.
Playing time tracks recent exposure in fifteen men's and women's football leagues: the denominator gradient and what it does to per-minute injury rates
preprint. Gustavo Pedro Ricou. Nicholas Mahony. Studies public appearance records from fifteen men's and women's football leagues to examine how exposure-dependent minutes can attenuate per-minute injury-rate associations. Separates outcome truncation from a denominator gradient; observational measurements do not establish causal injury predictions.
Report on Reflectometry
completed. Uses a controlled metal-plate reflector while Tore Supra was stopped to study reflectometry calibration, phase and Fourier analysis. A supervised signal-processing report, not a live-plasma experiment.
Disruption Studies on the COMPASS Tokamak
completed. Analyses magnetic diagnostics from the COMPASS tokamak using Python. The density target was not reached; island-mode identification and saddle-coil asymmetry remained inconclusive.
Automatic Data Control and Acquisition for the Study of Electrical Susceptibility and Thermally Stimulated Currents in Ceramics and Polarizable Films
completed. Integrates impedance, electrometer and temperature readings through LabVIEW instrument control and automatic logging for electrical-susceptibility and thermally stimulated current measurements. Supervised Physics Laboratory III work at the University of Porto.
Super-Precise Latency: How CPU Threading Affects High-Precision Latency, and an Industry-Wide Audit
in-preparation. Gustavo Pedro Ricou. Romaric Duvignau. Examines how CPU scheduling can distort sub-millisecond streaming-latency measurements, separating late clock readings from discarded samples. Controlled scheduling interventions and benchmark audits evaluate whether kernel timestamps exclude client-thread waits.