תקציר
We present an improved algorithm for properly learning convex polytopes in the realizable PAC setting from data with a margin. Our learning algorithm constructs a consistent polytope as an intersection of about t log t halfspaces with margins in time polynomial in t (where t is the number of halfspaces forming an optimal polytope). We also identify distinct generalizations of the notion of margin from hyperplanes to polytopes and investigate how they relate geometrically; this result may be of interest beyond the learning setting.
שפה מקורית | אנגלית |
---|---|
עמודים (מ-עד) | 5706-5716 |
מספר עמודים | 11 |
כתב עת | Advances in Neural Information Processing Systems |
כרך | 2018-December |
סטטוס פרסום | פורסם - 2018 |
אירוע | 32nd Conference on Neural Information Processing Systems, NeurIPS 2018 - Montreal, קנדה משך הזמן: 2 דצמ׳ 2018 → 8 דצמ׳ 2018 |