In this paper, we present a newviewof multiclass classification and introduce the constraint classification problem, a generalization that captures many flavors of multiclass classification. We provide the first optimal, distribution independent bounds for many multiclass learning algorithms, including winner-take-all (WTA). Based on our view, we present a learning algorithm that learns via a single linear classifier in high dimension. In addition to the distribution independent bounds, we provide a simple margin-based analysis improving generalization bounds for linear multiclass support vector machines.
|Original language||English (US)|
|Number of pages||15|
|Journal||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|State||Published - 2002|
ASJC Scopus subject areas
- Theoretical Computer Science
- Computer Science(all)