Viska Wei

Physics · In-Context Learning · Scalable ML

PhD Candidate in Physics; MSE in Computer Science and Applied Mathematics & Statistics
Johns Hopkins University

Advisor: Alexander S. Szalay, Vladimir Braverman, Fei Lu

My work connects stochastic processes, in-context learning theory, randomized algorithms, and machine learning for stellar spectroscopy, with one foot in mathematical analysis and the other in large-scale scientific data systems.

Featured

Current flagship papers and theory work.

Learning Interacting Particle Systems from Unlabeled Data

A trajectory-free method for learning interaction potentials from unlabeled particle snapshots via weak-form stochastic evolution equations

Stochastic ProcessesStatistical LearningInteracting Particle Systems

2026 · arXiv · Site · Code · Paper

SpecViT: Multi-Architecture Scaling Analysis for Stellar Parameter Estimation from PFS Medium-Resolution Spectra

A transformer-centered paper benchmark for stellar spectroscopy with architecture scaling, baseline comparisons, and transfer analysis

AstronomyTransformersSpectroscopy

2025 · Unpublished manuscript · Site · Code

The Generalization Dilemma in In-Context Learning: Task Diversity versus Estimation Accuracy

A phase-transition theory for when pretraining task diversity helps and when it hurts in-context inverse linear regression

In-Context LearningTransferPhase Transition

2026 · Working draft · Code

More Papers

Published papers, public preprints, and earlier work.

Denoising Stellar Spectra with U-Net Blindspot Neural Network

Self-supervised spectral denoising for low-SNR stellar observations with a blind-spot U-Net architecture

AstronomyDenoisingU-Net

2025 · Astronomy & Computing · Code

A UMAP-based Clustering Method for Multi-scale Damage Analysis of Laminates

UMAP geometry for discovering multi-scale damage patterns in composite laminate simulations

UMAPMaterialsScientific Computing

2022 · Applied Mathematical Modelling · Paper

Symmetric Norm Estimation and Regression on Sliding Windows

Sliding-window sketches for symmetric norm estimation and regression in streaming settings

StreamingAlgorithmsTheory

2021 · COCOON · Paper

Sketch and Scale: Geo-distributed tSNE and UMAP

A geo-distributed PCA → random sketching → UMAP pipeline for hundred-million-scale data exploration

Data ScienceDimensionality ReductionAlgorithms

2020 · IEEE BigData · Code · Paper

Bounding the Charm Yukawa Coupling

Supersymmetric Higgs phenomenology constraining the charm Yukawa coupling at the LHC

Particle PhysicsQFTCollider Phenomenology

2019 · Phys. Rev. D · Paper

Research Systems

Software and experimental platforms behind the papers.

Blade Agent

Multi-lane AI orchestrator for engineering and research workflows with quality gates

AI AgentOrchestrationPipeline

2025 · Site · Code