Oliver Hinder
University of Pittsburgh, PA·Computer Science
Training a machine learning model means running an optimization algorithm that adjusts the model step by step on data. These algorithms come with settings that must be tuned by hand for every new task.
- Fields
Computer Science
Mathematics & Statistics
- Topics
parameter-free optimization
stochastic optimization
machine learning
- Grants
- 1 known grant, through Apr 2027
What they work on
- Design stochastic optimization algorithms by first solving the noiseless problem and then extending them to noisy data with time-uniform concentration of measure results, instead of the standard online-to-batch conversion.
- Develop algorithms guided by trajectory cues, quantities observed during a run such as the distance between the current iterate and the starting point, that stand in for unknown problem parameters.
- Resolve outstanding theoretical problems in parameter-free optimization and turn the results into practical, open-source algorithms that reduce the burden of hand-tuning machine learning methods.
- Examine parameter-free optimization methods tailored to particular problem classes, including the training of deep learning models.
More about the research
Tuning takes a lot of computing power and skill, which shuts out many potential users, and even leading practitioners make costly tuning mistakes. This project aims to remove that burden by designing 'parameter-free' stochastic optimization methods. These are algorithms that learn from noisy, randomly sampled data and work well without being told key facts about the problem in advance.
Why it matters
Machine learning methods currently have to be tuned by hand for each new task, which demands large computing resources and expert skill, and mistakes are costly. Optimization algorithms that need no tuning could cut these costs and reduce errors. They could also open machine learning to users who lack specialist expertise or large computing budgets. The open-source software would make these methods directly available to practitioners.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Viswanath Nagarajan
University of Michigan, MI·Computer Science
Design approximation algorithms with provable performance guarantees for sequential testing of systems made of interconnected components, including series-parallel structures, under complex cost structures.
Nan Jiang
University of Illinois Urbana-Champaign, IL·Computer Science·Early-career
Develop model selection methods for offline reinforcement learning that choose among candidate policies using a holdout dataset, building on theoretical work on value-function selection.
Stephen Becker
University of Colorado Boulder, CO·Mathematics & Statistics
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.
The school
University of Pittsburgh: 83 labs on PI Pages, $1.5B a year on research, 403 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.