Stephen Becker
University of Colorado Boulder, CO·Mathematics & Statistics
Training and tuning artificial intelligence models usually comes down to optimization: searching for the settings that score best on some measure. Standard methods rely on gradients, the calculus-based slopes that point toward improvement.
- Fields
Mathematics & Statistics
Computer Science
Engineering
- Topics
gradient-free optimization
derivative-free optimization
zeroth-order optimization
- Grants
- 1 known grant, through Jun 2028
What they work on
- Develop gradient-free optimization algorithms for objective functions whose matrix-valued gradients have low intrinsic dimensionality, such as sparsity, low rank, or sparsity-plus-low-rank, to break worst-case efficiency barriers.
- Integrate structured gradient estimation into modern optimizers such as Muon, which rely on matrix factorizations, projections and solving linear systems in the Fisher information matrix.
- Devise computational linear algebra techniques for performing these matrix operations in the gradient-free setting, and release them as openly available software.
More about the research
In some problems gradients cannot be computed, so methods must work from trial evaluations alone. This gradient-free setting matters for small businesses, academic groups and public agencies that lack large computing infrastructure but still need to fine-tune machine learning models or optimize complex simulations. Gradient-free optimization has strong theoretical foundations but is poorly understood, and inefficient, when there are many variables.
Why it matters
Groups without large-scale computing, including small businesses, academic labs and public-sector organizations, could fine-tune machine learning models and optimize simulations far more cheaply. Making gradient-free optimization substantially more efficient in high dimensions would lower the computational barrier to entry for using modern AI methods.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Yuhua Zhu
University of California-Los Angeles, CA·Mathematics & Statistics
Develop a consensus-based interacting particle method, where many searchers share information to agree on the best point, for nonconvex optimization problems with equality constraints.
Michal Derezinski
University of Michigan, MI·Computer Science·Early-career
Reformulate optimal matrix sketching, the construction of a small randomized approximation of input data, using black-box sampling methods, so that sketch and sub-sample outputs come with reliability guarantees suitable for software.
Anna Ma
University of California-Irvine, CA·Mathematics & Statistics
Analyze the sample complexity of low-rank tensor completion, meaning how many observed entries are needed to recover a multi-dimensional dataset with simple underlying structure.
The school
University of Colorado Boulder: 36 labs on PI Pages, $718M a year on research, 447 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.