Convex Learning with Invariances
Choon Hui Teo, Amir Globerson, Sam Roweis, Alex Smola Poster ID: M38 PROBLEM: Insufficient amount of data to learn a robust classifier

Image invariance: translation, rotation, scaling

Text invariance: synonyms / vocabulary changes

Why: Want to use prior knowledge to improve classifier performance. What: Incorporate invariances into the loss function of a convex optimization problem. How: Employ well-studied and efficient optimization algorithms for learning.