Robert Pearce

Relevant coursework: Parallel Programming & Performance Engineering, NLP: Self-Supervised Models, ML: Deep Learning, ML: Coping with Non-Stationary Environments and Errors

Awards: Dean’s Honor List Medallion

Relevant coursework: Machine Learning, Computational Linear Algebra, Statistics, Operating Systems, Compilers, Databases, Cloud Computing.

Awards: Nominated for Outstanding Student in Mathematics

• Ported the kSZ zreion cosmological simulation code from CPU-based OpenMP to CUDA for GPU clusters, modernizing the Fortran codebase originally written for Bridges-2.
• Profiled the existing OpenMP routines to identify bottlenecks and prioritize the kernels with the most to gain from GPU acceleration.

• Automated 1,000 kSZ reionization simulations (~13,200 core-hours) on the Bridges-2 supercomputer using SLURM job arrays, sweeping a four-dimensional parameter space by Latin hypercube sampling at 128 cores per job.
• Built the data pipeline converting raw simulation output into ML-ready datasets: schema-validated HDF5 ingestion cutting storage 75%, and a flat-sky power spectrum estimator processing 1,000 maps in 3 minutes.
• Trained a PyTorch emulator mapping four reionization parameters to the five-bin kSZ angular power spectrum, reaching 5.12% MAPE on 200 held-out simulations, with the 625-parameter architecture informed by a Ray Tune ASHA search on the deterministic baseline.
• Quantified predictive uncertainty with Monte Carlo dropout at 201 passes per prediction, validating calibration against nominal coverage (67.1 / 89.7 / 97.1% at 1σ / 2σ / 3σ) and a reduced χ² of 1.50.
• Replaced the simulator inside a Bayesian likelihood, cutting a 960,000-evaluation emcee posterior from an infeasible 12.7M core-hours to 37 minutes on one CPU, validated to 1.2% on Thomson optical depth.
• Released the pipeline as reionemu, an open-source Python package on PyPI with a Zenodo DOI, GitHub Actions CI across Python 3.10–3.13, and a nine-page documentation site. Three versions, 141 commits, sole author.

• Selected from a competitive applicant pool to collaborate with a four-person team across three U.S. and South Korean universities, delivering a computational pathology study in four weeks (GitHub (opens in a new tab)).
• Built an end-to-end pipeline in Python and PyTorch, covering WSI segmentation, patching, and frozen UNI v2 feature extraction through attention-based MIL and Cox survival modeling, over 875 TCGA glioma patients (1,700+ slides) on a SLURM GPU cluster.
• Benchmarked MC-Dropout, Deep Ensembles, and SNGP for uncertainty quantification under 5-fold patient-level cross-validation; all three tied on discrimination (C-index 0.76), while SNGP delivered the most informative uncertainty (selective-prediction area 0.809) and best calibration.
• Achieved SNGP's calibrated uncertainty in a single forward pass, versus 100 stochastic passes for MC-Dropout and 5 separately trained models for Deep Ensembles, cutting inference cost substantially at equal accuracy.
• Diagnosed and fixed a per-fold risk-scaling artifact that had made SNGP appear worst-in-class, then built confound diagnostics isolating how much of each method's uncertainty was independent of its own risk score.

Identified and resolved 8 accessibility issues in iOS components, addressing issues related to WCAG 2.1 AA accessibility guidelines and enhancing usability for visually impaired users.

3+ years of teaching across multiple roles: SI Math Leader, Math Peer Mentor, Code Coach, and STEM Mentor. Supported 80+ students per semester in Calculus, Linear Algebra, and Statistics. Coached game development in Unity/Godot (C#, GDScript) and led robotics workshops for 200+ students using Sphero and Arduino platforms.

• Architected and published a pip-installable Python package for end-to-end scientific ML workflows: data ingestion, feature engineering, model training, and evaluation.
• Developed reusable pipelines for HDF5 data reduction and power spectrum computation, compressing intermediate data by 75% for large-scale experiments.
• Implemented a neural network emulator with Bayesian uncertainty quantification, achieving 5.1% error on held-out simulations.
• Integrated hyperparameter optimization via Ray Tune and CI/CD through GitHub Actions across Python 3.10–3.13, with versioned PyPI releases and a Zenodo DOI for reproducible experimentation.

• Uncertainty benchmark (MC-Dropout vs. Deep Ensembles vs. SNGP) for an ABMIL + Cox survival model on TCGA glioma whole-slide images, built on frozen UNI v2 features with 5-fold patient-level cross-validation.
• SNGP matched the others on discrimination (C-index 0.76) while giving the most informative uncertainty and best calibration in a single forward pass.