Learning theory & statistical inference
Model capacity, generalization, sample complexity, spectral testing, and identifiability.
Current roles
Yi-Shan (Eason) ChuResearch Assistant, Institute of Statistical Science, Academia Sinica · Independent AI Research Contractor, Taiwan Space Agency (TASA)Machine learning research grounded in mathematics, statistics, and real data.
My main research interests are theoretical machine learning, statistical learning, spectral methods, and geometric views of neural networks. I also work on remote sensing, computer vision, UAV localization, edge AI, and scientific computing.
Theoretical MLSpectral MethodsNeural GeometryRemote SensingEdge AI
ResearchLearning theory · spectral inference · neural geometry
Applied AIRemote sensing · UAV vision · edge deployment
Based inTaipei, Taiwan
Selected work
Selected projects and ongoing research.
LEARNING THEORY · 2025–2026
Learning Capacity of Kolmogorov–Arnold Networks
An NSTC-funded study of KAN model capacity, generalization, and sample complexity.
STATISTICS · 2026–Signal Dimension Estimation
Current work on eigenstructure-based tests and finite-sample identifiability.
NEURAL GEOMETRY · 2026S-GAI
A geometry-aware initialization method for sigmoidal MLPs; the corresponding work was accepted at ACCV 2026.
REMOTE SENSING · 2024–2026Post-disaster Segmentation
Transformer-based satellite-image segmentation developed from a TASA internship and later published in TAO.
COMPUTER VISION · 2025–2026GPS-Free UAV Visual Geo-localization
A vision-based localization workflow that matches UAV observations with satellite imagery to estimate position without relying on GPS.
Research interests
What I am working on now.
My projects span both theory and applications. Across them, I am particularly interested in what can be learned from finite data, how structure can be identified or used in models, and how methods behave under practical constraints.
Geometry of neural networks
Spectral and geometric structure for understanding representation, architecture, and initialization.
Applied AI systems
Remote sensing, UAV imagery, edge computing, computer vision, and deployment-oriented model development.
Publications & honors
Recent academic outputs and recognition.
Our paper "From Dataset Spectral Geometry to Network Weights: A Geometry-Aware Initialization for Sigmoidal MLPs in Image Classification" is accepted by ACCV 2026.
First / corresponding-author journal paper on post-disaster segmentation ‘Post-disaster affected area segmentation with vision transformer (ViT)-based model using Sentinel-2 and Formosat-5 imagery‘.
First Prize, System-Holdings Student Prize (Best Student Paper).
Research in practice
A few moments from presentations and project work.
Contact