Top-k Multiclass SVM

Spotlight | Paper | Long version | libsdca

  title = {Top-k Multiclass {SVM}},
  author = {Maksim Lapin and Matthias Hein and Bernt Schiele},
  booktitle = {NIPS},
  year = {2015}

Top-k SVM addresses the class ambiguity challenge

  • Ambiguous classes arise due to various reasons:
    • certain images exhibit multi-label nature;
    • in fine-grained classification it is by design;
    • large scale problems come along with non-negligible class overlap as well as label noise.
  • Top-k performance addresses this issue:
    • one guess is too hard even for humans;
    • allowing a few guesses is natural when the number of classes is large;
    • top-k error is already reported routinely in popular benchmarks, such as the ImageNet challenge.
  • In this work, we propose our Top-k Multiclass SVM, which
    • generalizes the Multiclass SVM of Crammer and Singer;
    • optimizes a tight convex upper bound on top-k error;
    • is trained efficiently via Stochastic Dual Coordinate Ascent (SDCA);
    • scales to large datasets such as ImageNet and Places;
    • demonstrates consistent improvements in top-k accuracy.