Geometry-aware learning
Local graph and kernel methods that adapt to heterogeneous density and manifold structure.
Ph.D. Candidate · Electrical & Computer Engineering · USC
Geometry, learning, and signal processing for structured data.
I develop machine learning methods that use geometry, spectral structure, and local information to make graphs, kernels, and learned representations more useful.
Graph learningKernel methodsRepresentation geometrySignal processing
Research
The thread connecting the work
Across graph learning, foundation-model embeddings, and signal analysis, the recurring question is simple: what structure is actually present in the data, and how should an algorithm adapt to it?
Spectral structure
Using effective rank and related spectral quantities to diagnose representation quality and graph behavior.
Foundation-model embeddings
Studying when learned embeddings preserve useful geometry for retrieval, fusion, and downstream inference.
Signal & image processing
Methods grounded in physical and statistical structure, from medical imaging to time-series signals.
Featured research story
Adaptive spectral bandwidth control
IEEE MLSP 2026 · with Antonio Ortega
Finding the scale where the kernel sees the right geometry.
This work proposes a per-node bandwidth criterion that matches the effective rank of a Gaussian-kernel operator to local intrinsic dimension, while anchoring the search in a manifold-consistent log-log regime. On CIFAR-100 SSL embeddings from six encoders, the adaptive approach improves graph-based inference over fixed and competing adaptive bandwidth choices.
Selected publications
Research output
IEEE MLSP 2026 · Accepted
Geometry-Aware Graph Construction via Adaptive Spectral Bandwidth Control
Per-node Gaussian bandwidth selection by matching kernel effective rank to local intrinsic dimension. Evaluated on six SSL encoders and downstream graph-based inference tasks.
Foundation models
Robust Classification under Noisy Labels: A Geometry-Aware Reliability Framework for Foundation Models
A geometry-aware reliability framework using local NNK neighborhoods for robust inference under label noise, evaluated on CIFAR-10 and DermaMNIST.
Signal processing
Non-Negative Kernel Graphs for Time-Varying Signals Using Visibility Graphs
An NNK graph formulation for time-varying signals with explicit temporal visibility, connecting sparse graph construction with signal structure.
Medical imaging
Magnetic Particle Imaging: SNR-Optimized Reconstruction & Relaxometry
Earlier work on signal-to-noise-optimized reconstruction and excitation-field effects in magnetic particle imaging, linking physical acquisition constraints to reconstruction quality.
Blog
Research, explained
Research note
Why Gaussian-kernel bandwidth has a useful middle
An intuitive explanation of the two failure regimes and why adaptive spectral control targets the middle.
Read post →Coming soon
From paper to research story
How to turn a technical contribution into a clear visual, a project page, and a useful public explanation.
DraftingComing soon
What graph geometry tells us about embeddings
Notes on local neighborhoods, kernel choice, and why representation quality is more than a single global score.
DraftingTalks & media
Explain it twice: once in the paper, once for humans.
YouTube talk slotDrop in a real YouTube embed when the recording is live.
Conference presentation
Designed for MLSP, Asilomar, seminars, and research explainers.
Podcast / interview slotAdd Spotify, Apple Podcasts, or RSS embed later.
Podcast & interviews
A home for longer-form conversations about research, AI, and building things.
Background
Experience & education
Ph.D.
University of Southern California
Electrical & Computer Engineering · Signal and Image Processing Institute.
Research
Machine learning, graph signal processing & imaging
Work spanning representation geometry, kernel graphs, noisy-label robustness, and medical imaging.
Earlier
Magnetic Particle Imaging
Research on SNR-aware reconstruction and relaxometry, with peer-reviewed outputs and conference presentations.
Contact
Research, collaboration, speaking.
For research collaborations, technical conversations, talks, or project questions, reach out directly.