Speaker: Nebojsa Jojic, Microsoft Research
Title: Counting Grid Models
A counting grid is an n-D grid of sparse distributions over features. A collection of dozens or more of such microtopics constitute constitutive elements in (ad)mixture models. These collections are limited to windows into a grid. For example, in a 64X64 grid we may only consider 10X10 windows into the grid, thus making each (ad)mixture component a collection of 100 microtopics, each associated to a location within the grid covered by the window. Two overlapping windows induce components with highly similar content, and gradual movement of the window across the grid evolves the resulting components slowly, thus allowing rather large grids to be trained even on smaller datasets without overtraining, as long as the window size is sufficiently large. I will discuss recent developments, such as the application of these models to images, local minima and training regimes, as well as the potential for deep architectures resulting from stacking these models.
For additional information contact: Jake Snell