Multiclass Learning from Contradictions

We introduce the notion of learning from contradictions, a.k.a We show that learning from contradictions (using MU-SVM) incurs lower sample complexity compared to multiclass SVM (M-SVM) by deriving the Natarajan dimension for sample complexity for PAC-learnability of MU-SVM. We also propose an analytic span bound for MU-SVM and demonstrate its utility for model selection resulting in $\sim 2-4 \times$ faster computation times than standard resampling techniques. We empirically demonstrate the efficacy of MU- SVM on several real world datasets achieving $ $ 20\% improvement in test accuracies compared to M-SVM. Insights into the underlying behavior of MU-SVM using a histograms-of-projections method are also provided.