The UCSD Zheng Lab focuses on scalable machine learning systems, designing algorithms and infrastructure for large datasets. Researchers emphasize reproducible workflows and open science tools that connect theory with real world deployment.
Through partnerships across campus, the group aligns its projects with translational impact, aiming for methods that support decision making under uncertainty and noisy real conditions.
| Name | Role | Expertise | Contact | Active Projects |
|---|---|---|---|---|
| Yibo Zheng | Principal Investigator | Large scale learning, optimization | zheng@ucsd.edu | Resource efficient training |
| Alice Carter | Senior Postdoc | Probabilistic modeling | acarter@ucsd.edu | Uncertainty calibration |
| Dev Patel | PhD Student | Streaming algorithms | dp@ucsd.edu | Edge inference |
| Jordan Lee | Research Engineer | Distributed systems | jlee@ucsd.edu | Cluster orchestration |
Scalable Learning Algorithms
Within this theme, the UCSD Zheng Lab develops first order and second order methods that scale to billions of examples. Work includes variance reduced SGD, adaptive quantization, and streaming formulations tailored for modern hardware.
Convergence Guarantees
The group provides nonasymptotic bounds that link algorithmic choices to generalization error, emphasizing how data geometry interacts with step size schedules.
Communication Efficient Training
Techniques such as gradient compression and delayed updates reduce bandwidth while preserving statistical performance, enabling cost effective training across data centers.
Systems For Real World Data
Another pillar of the UCSD Zheng Lab is building systems that handle messy, heterogeneous data from sensors, logs, and clinical records. Emphasis is placed on pipelines that integrate cleaning, feature extraction, and modeling into unified frameworks.
Data Ingestion And Versioning
Researchers design storage layouts and metadata standards that keep training sets traceable and comparable across experiments.
Deployment At Edge And Cloud
Solutions span low latency inference on edge devices and elastic serving in cloud environments, allowing models to move where bandwidth and latency constraints dictate.
Optimization Under Uncertainty
The lab studies how optimization objectives should change when measurements are noisy, censored, or generated by adaptive opponents. This work connects to bandits, robust optimization, and decision theory.
Robust Regularization
Instead of fitting to point estimates, models incorporate uncertainty sets that guard against distribution shift and label noise during training.
Sequential Decision Frameworks
Methods from online convex analysis are adapted to large scale settings, enabling policies that learn while controlling regret in realistic operational environments.
Open Science And Reproducibility
Members of the UCSD Zheng Lab release code, benchmarks, and detailed logs to support external verification. By publishing negative results and detailed configurations, they aim to narrow the gap between published results and deployed performance.
Benchmark Suite Construction
Carefully designed tasks isolate variables such as sample size, feature correlation, and label imbalance to support fair comparison across methods.
Reproducible Workflow Tooling
Containerized experiments, deterministic random seeds, and automated report generation allow collaborators and reviewers to trace every modeling choice.
Engagement And Impact
The UCSD Zheng Lab pursues work that translates algorithmic advances into operational systems with measurable benefits for users and institutions.
- Design scalable learning methods that respect computational and communication constraints
- Build systems that bridge research prototypes and production services
- Publish open benchmarks and tools to accelerate community progress
- Collaborate with domain experts to ensure models address real needs
- Emphasize uncertainty calibration and robustness in decision oriented models
FAQ
Reader questions
What types of real world problems does the UCSD Zheng Lab typically address?
The group targets problems in streaming prediction, resource efficient training, and uncertainty aware modeling for domains such as healthcare, sensor networks, and large scale web services.
How does the lab handle data privacy and sensitive information in its projects?
Researchers apply differential privacy, secure aggregation, and access controlled environments, aligning experimental designs with institutional review board requirements and industry standards.
Can external collaborators join ongoing projects led by the UCSD Zheng Lab?
Yes, the lab welcomes collaborators with complementary data, domain expertise, or deployment channels, coordinating through joint grant proposals and shared milestone plans.
What open resources does the group provide to support reproducibility?
They release benchmark datasets, reference implementations, and detailed experiment metadata, enabling independent verification and facilitating reuse across research teams.