Timezone: »

Learning Structural Sparsity
Yoram Singer

Thu Dec 09 04:00 PM -- 04:30 PM (PST) @

In the past years my work focused on algorithms for learning high dimensional yet sparse models from very large datasets. During the two years that Sam spent at Google, he greatly influenced my course of research on large scale learning of structural sparsity. He was too humble and too busy to formally co-author any of the papers that constitute the talk (see http://magicbroom.info/Sparsity.html). Yet, many parts of this talk would not have materialized without his encouragement, feedback, and ideas. In the talk I review the design, analysis and implementation of sparsity promoting learning algorithms, including coordinate and mirror descent with non-smooth regularization, forward-backward splitting algorithms, and other recently devised algorithms for sparse models. I will conclude with an overview of new work on learning self pruning decision trees and structured histograms by combining exponential models with sparsity promoting regularization.

Author Information

Yoram Singer (Princeton)

More from the Same Authors