About
Hello, I am Joseph Dehoney, but I mostly go by Joe (or Joda depending on when we met).
I'm a Stanford M.S. CS student with a machine learning background spanning theory, systems, and applications.
I enjoy building practical ML systems end-to-end while grounding decisions in principled learning objectives.
I also have a research background in quantum information theory and quantum complexity,
with an emphasis on mathematically rigorous analysis of computational limits, verification, and learnability in structured physical systems.
I am especially interested in the intersection of ML and quantum: where learning methods
meet information-theoretic limits, and where structure can turn hard scientific problems into tractable
algorithms.
Research interests
The main directions I am currently pursuing are:Machine Learning (Theory, Systems, Applications)
- Learning theory perspective: emphasizing identifiability, generalization, and complexity-aware modeling choices.
- Robust ML systems: building reproducible data/model pipelines with strong validation and monitoring.
- Scalable training: distributed and memory-aware workflows for large multimodal and graph data.
- Applied scientific ML: using machine learning for biomedical and neuroimaging workflows.
Quantum Information Theory Research
- Verification & structure: when does mathematical structure or physical constraints allow for quantum advantage?
- Learning in many-body settings: identifiability and complexity of learning local Hamiltonians / Gibbs states from local data.
- Complexity boundaries: mapping efficient methods versus provably hard quantum tasks.
Intersection: ML × Quantum
- Learning under physical constraints: designing estimation procedures that respect locality, noise, and measurement limits.
- Theory to practice: connecting information-theoretic guarantees to implementable algorithms.
- Scientific tooling: building reliable computational methods for physics and engineering research.
Recent work
Quantum Information Theory
Towards QCMA containment for Kronecker coefficient positivity
Investigating whether hard instances of Kronecker positivity admit efficiently checkable witnesses; includes extending a known hardness result for certain 2-point correlation functions to a 3-point setting (with Adam Bouland & Tamara Kohler).
Quantum optimization of MIS with Rydberg atom arrays (course presentation)
Synthesized a paper on solving Max Independent Set with programmable Rydberg-atom arrays, focusing on the MIS to blockade mapping and how control schedules (detuning/Rabi) realize analog optimization. Presented the work as a course final project in quantum control and engineering.
Machine Learning
Scaling PyG: Taming Massive Graphs (article + colab)
Built a memory-safe GNN training workflow with neighborhood/hierarchical sampling and distributed loaders in PyTorch Geometric, with reproducible throughput and accuracy tradeoff benchmarks.
Intersection: ML × Quantum
Learning Quantum Gibbs States Locally & Efficiently
Combined quantum many-body structure with learning-theoretic analysis to characterize when local thermal data is sufficient for efficient Hamiltonian recovery.
Background
I build data pipelines at Stanford for multimodal biomedical and neuroimaging research datasets (automation, CI validation, Slurm/Docker/AWS workflows). I enjoy the full lifecycle: from idea → implementation → results → writeup. Further details on the outcomes of this work can be seen in my publications section.