Qi Yu
University of California-San Diego, CA·Computer Science·Early-career
Deep learning powers many modern artificial intelligence (AI) systems. However, it often depends on trial and error, huge datasets and heavy computing, and it can fail on situations it has not seen before.
- Fields
Computer Science
Mathematics & Statistics
- Topics
deep learning
machine learning
spatiotemporal data
- Grants
- 2 known grants, through Jun 2027
What they work on
- Develop deep learning models that forecast spatiotemporal dynamics while obeying physical laws, with the goal of simulating ocean currents, traffic and epidemic spread faster than numerical solvers.
- Create methods that infer complex interactions in large spatiotemporal datasets and quantify the uncertainty of forecasts so they can support real-time decisions.
- Study the trade-off between symmetry in a model and symmetry in the data, using neural networks that combine equivariant and non-equivariant features, grounded in representation theory.
- Learn symmetries directly from data, and use quiver representations to describe the symmetry of deep network parameter spaces, to improve how models generalize.
More about the research
It also struggles with spatiotemporal data, meaning data collected across space and time, such as ocean currents or the spread of an epidemic. This kind of data changes quickly and follows physical laws. The research combines two directions: building deep learning tools for such data, and developing mathematical theory, based on symmetry, that explains how deep learning works.
Why it matters
Fast AI models that can be trusted to obey physical laws could allow real-time planning for systems such as ocean currents, traffic and epidemics. These models could support decisions in weather science and public health. A firmer theory of symmetry in deep learning could lead to models that use scientific knowledge, offer more reliable performance, and adapt to new situations with less data.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
George Bebis
University of Nevada-Reno, NV·Computer Science
Develop a physics-informed conditional generative deep learning framework that infers underground fracture network geometry and hydraulic properties from hydraulic and tracer tomography data.
Chao Zhang
Georgia Institute of Technology, GA·Computer Science·Early-career
Develop deep predictive models for spatial networks that capture how they change over space and time while estimating the uncertainty of their own forecasts.
Susanta Ghosh
Michigan Technological University, MI·Engineering·Early-career
Develop a Bayesian convolutional neural network that predicts electron density, the map of where electrons reside in a material, while reporting its own uncertainty.
The school
University of California-San Diego: 84 labs on PI Pages, $1.9B a year on research, 637 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.