Collecting labeled data is costly and thus a critical bottleneck in
real-world classification tasks. To mitigate this problem, we propose a novel
setting, namely learning from complementary labels for multi-class
classification. A complementary label specifies a class that a pattern do