A random-class model for evaluating highly multi-class classifcation tasks.

Seminar | April 12 | 4-5 p.m. | 1011 Evans Hall

Speaker/Performer:  Yuval Benjamini, Hebrew University of Jerusalem

Sponsor:  Department of Statistics

In many multi-class classification tasks, the potential set of classes is vast (think face detection or speaker identification).
Researchers often design and evaluate the classifiers on a subset from the population of classes to which they will be applied.
In this talk, I argue that it may be useful to model the observed set of classes as a random sample from a population.
As the main results, I will present parametric and nonparametric characterizations for the
effect of the number of classes on classification accuracy in a few-shot learning design.
For the non-parametric model, which empirically better describes real data-sets, we develop practical methods for
estimation with relatively few observed classes. This allows us to predict the classification accuracy as more classes are added.
I also discuss use-cases from neuroscience use-cases including evaluating representations in brain-decoding tasks and subject privacy.
The talk is based on work with Yuli Slavutsky and Charles Zheng.

Event Contact:  CA, zhivotovskiy@berkeley.edu, 5102292370

Access Coordinator:  Nikita Zhivotovskiy,  zhivotovskiy@berkeley.edu,  510-229-2370