Special Colloquium
Clément Canonne
IBM Research Almaden
Data Incognita: Inference in Uncharted Lands
Abstract: The vast amounts of data now available, and our increasing reliance on
it, have come to the front and center of machine learning, computer
science, permeating numerous aspects of society. The need to process and
analyze this data has raised critical algorithmic and
information-theoretic challenges, as well as pressing societal concerns.
In this talk, I will present two research directions aimed at addressing
components of these challenges. In the first part of this talk, I will
present a general framework for data inference under "local information
constraints," which captures a range of scenarios such as distributed
inference under stringent privacy or communication requirements.
The second reconsiders the data access model itself: I introduce new
models of data collection, that capture key features in many practical
settings. I show that these models allow for significantly more
efficient testing and estimation, in comparison to traditional settings,
e.g., assuming access to i.i.d. observations.
Wednesday February 19, 2020 at 3:00 PM in 636 SEO